gem_id
stringlengths
37
41
paper_id
stringlengths
3
4
paper_title
stringlengths
19
183
paper_abstract
stringlengths
168
1.38k
paper_content
dict
paper_headers
dict
slide_id
stringlengths
37
41
slide_title
stringlengths
2
85
slide_content_text
stringlengths
11
2.55k
target
stringlengths
11
2.55k
references
list
GEM-SciDuet-train-8#paper-972#slide-9
972
Obtaining SMT dictionaries for related languages
This study explores methods for developing Machine Translation dictionaries on the basis of word frequency lists coming from comparable corpora. We investigate (1) various methods to measure the similarity of cognates between related languages, (2) detection and removal of noisy cognate translations using SVM ranking. ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", ...
GEM-SciDuet-train-8#paper-972#slide-9
Manual evaluation
Conclusions Results Machine Translation Results on sample of 100 words n-best lists L, L-R, L-C ranking model E List L List L-R List L-C
Conclusions Results Machine Translation Results on sample of 100 words n-best lists L, L-R, L-C ranking model E List L List L-R List L-C
[]
GEM-SciDuet-train-8#paper-972#slide-10
972
Obtaining SMT dictionaries for related languages
This study explores methods for developing Machine Translation dictionaries on the basis of word frequency lists coming from comparable corpora. We investigate (1) various methods to measure the similarity of cognates between related languages, (2) detection and removal of noisy cognate translations using SVM ranking. ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", ...
GEM-SciDuet-train-8#paper-972#slide-10
Addition of lists SMT
1-best lists with L-C and E ranking pt-es: 80K training sentences, 100K cognate pairs significant uk-ru: 140K training sentences, 100K cognate pairs
1-best lists with L-C and E ranking pt-es: 80K training sentences, 100K cognate pairs significant uk-ru: 140K training sentences, 100K cognate pairs
[]
GEM-SciDuet-train-8#paper-972#slide-12
972
Obtaining SMT dictionaries for related languages
This study explores methods for developing Machine Translation dictionaries on the basis of word frequency lists coming from comparable corpora. We investigate (1) various methods to measure the similarity of cognates between related languages, (2) detection and removal of noisy cognate translations using SVM ranking. ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", ...
GEM-SciDuet-train-8#paper-972#slide-12
Conclusions
MT dictionaries extracted from comparable resources for related languages Positive results on the n-bes lists with L-C Frequency window heuristic shows poor results ML models are able to rank similar words on the top of the list Preliminary results on an SMT system show modest improvements compare to the baseline The O...
MT dictionaries extracted from comparable resources for related languages Positive results on the n-bes lists with L-C Frequency window heuristic shows poor results ML models are able to rank similar words on the top of the list Preliminary results on an SMT system show modest improvements compare to the baseline The O...
[]
GEM-SciDuet-train-8#paper-972#slide-13
972
Obtaining SMT dictionaries for related languages
This study explores methods for developing Machine Translation dictionaries on the basis of word frequency lists coming from comparable corpora. We investigate (1) various methods to measure the similarity of cognates between related languages, (2) detection and removal of noisy cognate translations using SVM ranking. ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", ...
GEM-SciDuet-train-8#paper-972#slide-13
Future work
Morphology features for the n-best list (Unsupervised) Instead of prefix heuristic (L-C) and stemmer (L-R) Contribution for all the produced cognate lists on SMT Using char-based transliteration model trained on Zoo plus n-best lists Motivation alignment learns useful transformations: e.g. introducao (pt) vs introducci...
Morphology features for the n-best list (Unsupervised) Instead of prefix heuristic (L-C) and stemmer (L-R) Contribution for all the produced cognate lists on SMT Using char-based transliteration model trained on Zoo plus n-best lists Motivation alignment learns useful transformations: e.g. introducao (pt) vs introducci...
[]
GEM-SciDuet-train-9#paper-975#slide-0
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-0
Latent Dirichlet Allocation
David Blei. Probabilistic topic models. Comm. ACM. 2012
David Blei. Probabilistic topic models. Comm. ACM. 2012
[]
GEM-SciDuet-train-9#paper-975#slide-2
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-2
Variations and extensions
Author topic model (Rosen-Zvi et al 2004) Supervised LDA (SLDA; McAuliffe and Blei, 2008) Dirichlet multinomial regression (Mimno and McCallum, 2008) Sparse additive generative models (SAGE; Eisenstein et al, Structural topic model (Roberts et al, 2014)
Author topic model (Rosen-Zvi et al 2004) Supervised LDA (SLDA; McAuliffe and Blei, 2008) Dirichlet multinomial regression (Mimno and McCallum, 2008) Sparse additive generative models (SAGE; Eisenstein et al, Structural topic model (Roberts et al, 2014)
[]
GEM-SciDuet-train-9#paper-975#slide-3
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-3
Desired features of model
Easy modification by end-users. Covariates: features which influences text (as in SAGE). Labels: features to be predicted along with text (as in SLDA). Possibility of sparse topics. Incorporate additional prior knowledge. Use variational autoencoder (VAE) style of inference (Kingma
Easy modification by end-users. Covariates: features which influences text (as in SAGE). Labels: features to be predicted along with text (as in SLDA). Possibility of sparse topics. Incorporate additional prior knowledge. Use variational autoencoder (VAE) style of inference (Kingma
[]
GEM-SciDuet-train-9#paper-975#slide-4
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-4
Desired outcome
Coherent groupings of words (something like topics), with offsets for observed metadata Encoder to map from documents to latent representations Classifier to predict labels from from latent representation
Coherent groupings of words (something like topics), with offsets for observed metadata Encoder to map from documents to latent representations Classifier to predict labels from from latent representation
[]
GEM-SciDuet-train-9#paper-975#slide-5
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-5
Model
p( w) i generator network: p(w i) = fg( ) ELBO Eq[log p(words ri DKL[q(ri words)p(ri encoder network: q( i w) = fe( )
p( w) i generator network: p(w i) = fg( ) ELBO Eq[log p(words ri DKL[q(ri words)p(ri encoder network: q( i w) = fe( )
[]
GEM-SciDuet-train-9#paper-975#slide-6
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-6
Scholar
p(word i ci softmax(d Ti B(topic) cTi B(cov)) Optionally include interactions between topics and covariates p(yi i ci fy (i ci log i f(words, ci yi Optional incorporation of word vectors to embed input
p(word i ci softmax(d Ti B(topic) cTi B(cov)) Optionally include interactions between topics and covariates p(yi i ci fy (i ci log i f(words, ci yi Optional incorporation of word vectors to embed input
[]
GEM-SciDuet-train-9#paper-975#slide-7
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-7
Optimization
Tricks from Srivastava and Sutton, 2017: Adam optimizer with high-learning rate to bypass mode collapse Batch-norm layers to avoid divergence Annealing away from batch-norm output to keep results interpretable
Tricks from Srivastava and Sutton, 2017: Adam optimizer with high-learning rate to bypass mode collapse Batch-norm layers to avoid divergence Annealing away from batch-norm output to keep results interpretable
[]
GEM-SciDuet-train-9#paper-975#slide-8
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-8
Output of Scholar
B(topic),B(cov): Coherent groupings of positive and negative deviations from background ( topics) f, f: Encoder network: mapping from words to topics: i softmax(fe(words, ci yi fy : Classifier mapping from i to labels: y fy (i ci
B(topic),B(cov): Coherent groupings of positive and negative deviations from background ( topics) f, f: Encoder network: mapping from words to topics: i softmax(fe(words, ci yi fy : Classifier mapping from i to labels: y fy (i ci
[]
GEM-SciDuet-train-9#paper-975#slide-9
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-9
Evaluation
1. Performance as a topic model, without metadata (perplexity, coherence) 2. Performance as a classifier, compared to SLDA 3. Exploratory data analysis
1. Performance as a topic model, without metadata (perplexity, coherence) 2. Performance as a classifier, compared to SLDA 3. Exploratory data analysis
[]
GEM-SciDuet-train-9#paper-975#slide-10
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-10
Quantitative results basic model
LDA SAGE NVDM Scholar Scholar Scholar +wv +sparsity
LDA SAGE NVDM Scholar Scholar Scholar +wv +sparsity
[]
GEM-SciDuet-train-9#paper-975#slide-11
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-11
Classification results
LR SLDA Scholar Scholar (labels) (covariates)
LR SLDA Scholar Scholar (labels) (covariates)
[]
GEM-SciDuet-train-9#paper-975#slide-12
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-12
Exploratory Data Analysis
Data: Media Frames Corpus (Card et al, 2015) Collection of thousands of news articles annotated in terms of tone and framing Relevant metadata: year of publication, newspaper, etc.
Data: Media Frames Corpus (Card et al, 2015) Collection of thousands of news articles annotated in terms of tone and framing Relevant metadata: year of publication, newspaper, etc.
[]
GEM-SciDuet-train-9#paper-975#slide-13
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-13
Tone as a label
english language city spanish community boat desert died men miles coast haitian visas visa applications students citizenship asylum judge appeals deportation court labor jobs workers percent study wages bush border president bill republicans state gov benefits arizona law bill bills arrested charged charges agents ope...
english language city spanish community boat desert died men miles coast haitian visas visa applications students citizenship asylum judge appeals deportation court labor jobs workers percent study wages bush border president bill republicans state gov benefits arizona law bill bills arrested charged charges agents ope...
[]
GEM-SciDuet-train-9#paper-975#slide-14
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-14
Tone as a covariate with interactions
Base topics Anti-immigration Pro-immigration ice customs agency population born percent judge case court guilty patrol border miles licenses drivers card island story chinese guest worker workers benefits bill welfare criminal customs jobs million illegals guilty charges man patrol border foreign sept visas smuggling f...
Base topics Anti-immigration Pro-immigration ice customs agency population born percent judge case court guilty patrol border miles licenses drivers card island story chinese guest worker workers benefits bill welfare criminal customs jobs million illegals guilty charges man patrol border foreign sept visas smuggling f...
[]
GEM-SciDuet-train-9#paper-975#slide-15
975
Neural Models for Documents with Metadata
Most real-world document collections involve various types of metadata, such as author, source, and date, and yet the most commonly-used approaches to modeling text corpora ignore this information. While specialized models have been developed for particular applications, few are widely used in practice, as customizatio...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsi...
GEM-SciDuet-train-9#paper-975#slide-15
Conclusions
Variational autoencoders (VAEs) provide a powerful framework for latent variable modeling We use the VAE framework to create a customizable model for documents with metadata We obtain comparable performance with enhanced flexibility and scalability Code is available: www.github.com/dallascard/scholar
Variational autoencoders (VAEs) provide a powerful framework for latent variable modeling We use the VAE framework to create a customizable model for documents with metadata We obtain comparable performance with enhanced flexibility and scalability Code is available: www.github.com/dallascard/scholar
[]
GEM-SciDuet-train-10#paper-977#slide-0
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-0
Language generation Equivalence in the target space
Ground truth sequences lie in a union of low-dimensional subspaces where sequences convey the same message. I France won the world cup for the second time. I France captured its second world cup title. Some words in the vocabulary share the same meaning. I Capture, conquer, win, gain, achieve, accomplish, . . . ACL 201...
Ground truth sequences lie in a union of low-dimensional subspaces where sequences convey the same message. I France won the world cup for the second time. I France captured its second world cup title. Some words in the vocabulary share the same meaning. I Capture, conquer, win, gain, achieve, accomplish, . . . ACL 201...
[]
GEM-SciDuet-train-10#paper-977#slide-1
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-1
Contributions
Take into consideration the nature of the target language space with: A token-level smoothing for a robust multi-class classification. A sequence-level smoothing to explore relevant alternative sequences. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
Take into consideration the nature of the target language space with: A token-level smoothing for a robust multi-class classification. A sequence-level smoothing to explore relevant alternative sequences. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-2
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-2
Maximum likelihood estimation MLE
For a pair (x y), we model the conditional distribution: Given the ground truth target sequence y?: ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Zero-one loss, all the outputs y y? are treated equally. Discrepancy at the sentence level between the training (1-gram) and evaluation metr...
For a pair (x y), we model the conditional distribution: Given the ground truth target sequence y?: ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Zero-one loss, all the outputs y y? are treated equally. Discrepancy at the sentence level between the training (1-gram) and evaluation metr...
[]
GEM-SciDuet-train-10#paper-977#slide-3
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-3
Loss smoothing
Prerequisite: A word embedding w (e.g. Glove) in the target space and a distance d with a temperature st. r
Prerequisite: A word embedding w (e.g. Glove) in the target space and a distance d with a temperature st. r
[]
GEM-SciDuet-train-10#paper-977#slide-4
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-4
Token level smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-5
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-5
Loss smoothing Token level
Uniform label smoothing over all words in the vocabulary: We can leverage word co-occurrence statistics to build a non-uniform and meaningful distribution. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing We can estimate the exact KL divergence for every target token.
Uniform label smoothing over all words in the vocabulary: We can leverage word co-occurrence statistics to build a non-uniform and meaningful distribution. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing We can estimate the exact KL divergence for every target token.
[]
GEM-SciDuet-train-10#paper-977#slide-6
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-6
Sequence level smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-7
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-7
Loss smoothing Sequence level
Prerequisite: A distance d in the sequences space Vn, n N. Hamming Edit 1BLEU 1CIDEr ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Can we evaluate the partition function Z for a given reward? We can approximate Z for Hamming distance.
Prerequisite: A distance d in the sequences space Vn, n N. Hamming Edit 1BLEU 1CIDEr ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Can we evaluate the partition function Z for a given reward? We can approximate Z for Hamming distance.
[]
GEM-SciDuet-train-10#paper-977#slide-8
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-8
Loss smoothing Sequence level Hamming distance
consider only sequences of the same length as y? (d(y y if |y |y We partition the set of sequences y?: their distance to the ground truth d d Sd Sd The reward in each subset is a constant. The cardinality of each subset is known. d Z |Sd exp ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothin...
consider only sequences of the same length as y? (d(y y if |y |y We partition the set of sequences y?: their distance to the ground truth d d Sd Sd The reward in each subset is a constant. The cardinality of each subset is known. d Z |Sd exp ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothin...
[]
GEM-SciDuet-train-10#paper-977#slide-9
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-9
Loss smoothing Sequence level Other distances
We cannot easily sample from more complicated rewards such as BLEU or CIDEr. Choose q the reward distribution relative to Hamming distance. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
We cannot easily sample from more complicated rewards such as BLEU or CIDEr. Choose q the reward distribution relative to Hamming distance. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-10
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-10
Loss smoothing Sequence level Support reduction
Can we reduce the support of r? Reduce the support from V |y?| to V |y sub where Vsub V. Vsub Vbatch tokens occuring in the SGD mini-batch. Vsub Vrefs tokens occuring in the available references. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
Can we reduce the support of r? Reduce the support from V |y?| to V |y sub where Vsub V. Vsub Vbatch tokens occuring in the SGD mini-batch. Vsub Vrefs tokens occuring in the available references. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-11
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-11
Loss smoothing Sequence level Lazy training
Default training Lazy training l y l is: l y l is: not forwarded in the RNN. log p(yl |yl x) ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing |y ||cell |, where cell are the cell parameters.
Default training Lazy training l y l is: l y l is: not forwarded in the RNN. log p(yl |yl x) ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing |y ||cell |, where cell are the cell parameters.
[]
GEM-SciDuet-train-10#paper-977#slide-12
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-12
Image captioning on MS COCO Setup
5 captions for every image. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
5 captions for every image. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-13
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-13
Image captioning on MS COCO Results
Loss Reward Vsub BLEU-1 BLEU-4 CIDEr ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
Loss Reward Vsub BLEU-1 BLEU-4 CIDEr ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-14
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-14
Machine translation Setup
Bi-LSTM encoder-decoder with attention (Bahdanau et al. 2015) IWSLT14 DEEN WMT14 ENFR Dev 7k Dev 6k Test 7k Test 3k ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
Bi-LSTM encoder-decoder with attention (Bahdanau et al. 2015) IWSLT14 DEEN WMT14 ENFR Dev 7k Dev 6k Test 7k Test 3k ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-15
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-15
Machine translation Results
Loss Reward Vsub WMT14 EnFr IWSLT14 DeEn ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
Loss Reward Vsub WMT14 EnFr IWSLT14 DeEn ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-16
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-16
Conclusion
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-17
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-17
Takeaways
Improving over MLE with: Sequence-level smoothing: an extension of RAML (Norouzi et al. 2016) I Reduced support of the reward distribution. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Token-level smoothing: smoothing across semantically similar tokens instead of the usual uniform noi...
Improving over MLE with: Sequence-level smoothing: an extension of RAML (Norouzi et al. 2016) I Reduced support of the reward distribution. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing Token-level smoothing: smoothing across semantically similar tokens instead of the usual uniform noi...
[]
GEM-SciDuet-train-10#paper-977#slide-18
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-18
Future work
Validate on other seq2seq models besides LSTM encoder-decoders. Validate on models with BPE instead of words. I Experiment with other distributions for sampling other than the Hamming distance. I Sparsify the reward distribution for scalability. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoo...
Validate on other seq2seq models besides LSTM encoder-decoders. Validate on models with BPE instead of words. I Experiment with other distributions for sampling other than the Hamming distance. I Sparsify the reward distribution for scalability. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoo...
[]
GEM-SciDuet-train-10#paper-977#slide-19
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-19
Appendices
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-20
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-20
Training time
Average wall time to process a single batch (10 images 50 captions) when training the RNN language model with fixed CNN (without attention) on a Titan X GPU. Loss MLE Tok Seq Seq lazy Seq Seq lazy Seq Seq lazy Tok-Seq Tok-Seq Tok-Seq Reward Glove sim Hamming Vsub V V Vbatch Vbatch Vrefs Vrefs V Vbatch Vrefs ACL 2018, M...
Average wall time to process a single batch (10 images 50 captions) when training the RNN language model with fixed CNN (without attention) on a Titan X GPU. Loss MLE Tok Seq Seq lazy Seq Seq lazy Seq Seq lazy Tok-Seq Tok-Seq Tok-Seq Reward Glove sim Hamming Vsub V V Vbatch Vbatch Vrefs Vrefs V Vbatch Vrefs ACL 2018, M...
[]
GEM-SciDuet-train-10#paper-977#slide-21
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-21
Generated captions
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoothing
[]
GEM-SciDuet-train-10#paper-977#slide-22
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-22
Generated translations EnFr
I think its conceivable that these data are used for mutual benefit. Jestime quil est concevable que ces donnees soient utilisees dans leur interet mutuel. Je pense quil est possible que ces donnees soient utilisees a des fins reciproques. Je pense quil est possible que ces donnees soient utilisees pour le benefice mut...
I think its conceivable that these data are used for mutual benefit. Jestime quil est concevable que ces donnees soient utilisees dans leur interet mutuel. Je pense quil est possible que ces donnees soient utilisees a des fins reciproques. Je pense quil est possible que ces donnees soient utilisees pour le benefice mut...
[]
GEM-SciDuet-train-10#paper-977#slide-23
977
Token-level and sequence-level loss smoothing for RNN language models
Despite the effectiveness of recurrent neural network language models, their maximum likelihood estimation suffers from two limitations. It treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. Second, it suffers from "exposure bias": during training tokens...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss s...
GEM-SciDuet-train-10#paper-977#slide-23
MS COCO server results
BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CIDEr SPICE Ours: Tok-Seq CIDEr Ours: Tok-Seq CIDEr + Table: MS-COCO s server evaluation . (+) for ensemble submissions, for submissions with CIDEr optimization and () for models using additional data. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoot...
BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CIDEr SPICE Ours: Tok-Seq CIDEr Ours: Tok-Seq CIDEr + Table: MS-COCO s server evaluation . (+) for ensemble submissions, for submissions with CIDEr optimization and () for models using additional data. ACL 2018, Melbourne M. Elbayad || Token-level and Sequence-level Loss Smoot...
[]
GEM-SciDuet-train-11#paper-978#slide-0
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-0
Background
based on predefined event schema and rich features encoded from annotated event Pros: extract high quality events for predefined types Cons: require large amount of human annotations and cannot extract event mentions for new event types Traditional Event Extraction Pipeline Consumer 1: I want an event extractor for Tra...
based on predefined event schema and rich features encoded from annotated event Pros: extract high quality events for predefined types Cons: require large amount of human annotations and cannot extract event mentions for new event types Traditional Event Extraction Pipeline Consumer 1: I want an event extractor for Tra...
[]
GEM-SciDuet-train-11#paper-978#slide-1
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-1
Motivation
Zero Shot Learning for Event Extraction both event mentions and types have rich semantics and structures, which can specify their consistency and connections E1. The Government of China has ruled Tibet since 1951 after dispatching troops to the E2. Iranian state television stated that the conflict between the Iranian p...
Zero Shot Learning for Event Extraction both event mentions and types have rich semantics and structures, which can specify their consistency and connections E1. The Government of China has ruled Tibet since 1951 after dispatching troops to the E2. Iranian state television stated that the conflict between the Iranian p...
[]
GEM-SciDuet-train-11#paper-978#slide-3
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-3
Approach Details
Trigger and Argument Identification AMR parsing and FrameNet verbs/nominal lexical units Subset of AMR relations None-Core Roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Event and Type Structure Construction Structu...
Trigger and Argument Identification AMR parsing and FrameNet verbs/nominal lexical units Subset of AMR relations None-Core Roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Event and Type Structure Construction Structu...
[]
GEM-SciDuet-train-11#paper-978#slide-4
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-4
Evaluation
Zero-Shot Classification for ACE Events Given trigger and argument boundaries, use a subset of ACE types for training, and remained types for testing Seen types for each experiment setting Setting Top-N Seen Types for Training/Dev D Attack, Transport, Die, Meet, Arrest-Jail, Transfer-Money, Sentence, Elect, Transfer-Ow...
Zero-Shot Classification for ACE Events Given trigger and argument boundaries, use a subset of ACE types for training, and remained types for testing Seen types for each experiment setting Setting Top-N Seen Types for Training/Dev D Attack, Transport, Die, Meet, Arrest-Jail, Transfer-Money, Sentence, Elect, Transfer-Ow...
[]
GEM-SciDuet-train-11#paper-978#slide-5
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-5
Discussion
Impact of AMR Parsing AMR is used to identify candidate triggers and arguments, as well as construct event structures Compare AMR with Semantic Role Labeling (SRL) on a subset of ERE corpus with perfect AMR annotations Train on top-6 most popular seen (training) types: Arrest-Jail, Execute, Die, Meet, Sentence, Charge-...
Impact of AMR Parsing AMR is used to identify candidate triggers and arguments, as well as construct event structures Compare AMR with Semantic Role Labeling (SRL) on a subset of ERE corpus with perfect AMR annotations Train on top-6 most popular seen (training) types: Arrest-Jail, Execute, Die, Meet, Sentence, Charge-...
[]
GEM-SciDuet-train-11#paper-978#slide-6
978
Zero-Shot Transfer Learning for Event Extraction
Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "...
GEM-SciDuet-train-11#paper-978#slide-6
Conclusion and Future Work
We model event extraction as a generic grounding problem, instead of classification By leveraging existing human constructed event schemas and manual annotations for a small set of seen types, the zero shot framework can improve the scalability of event extraction and save human effort In the future, we will extend thi...
We model event extraction as a generic grounding problem, instead of classification By leveraging existing human constructed event schemas and manual annotations for a small set of seen types, the zero shot framework can improve the scalability of event extraction and save human effort In the future, we will extend thi...
[]
GEM-SciDuet-train-12#paper-980#slide-0
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-0
An Example Dialogue with Movie Bot
Actual dialogues can be more complex: Speech/Natural language understanding errors o Input may be spoken language form o Need to reason under uncertainty o Revise information collected earlier Source code available at https://github/com/MiuLab/TC-Bot
Actual dialogues can be more complex: Speech/Natural language understanding errors o Input may be spoken language form o Need to reason under uncertainty o Revise information collected earlier Source code available at https://github/com/MiuLab/TC-Bot
[]
GEM-SciDuet-train-12#paper-980#slide-1
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-1
Task oriented slot filling Dialogues
Domain: movie, restaurant, flight, Slot: information to be filled in before completing a task o For Movie-Bot: movie-name, theater, number-of-tickets, price, o Inspired by speech act theory (communication as action) request, confirm, inform, thank-you, o Some may take parameters: "Is Kungfu Panda the movie you are look...
Domain: movie, restaurant, flight, Slot: information to be filled in before completing a task o For Movie-Bot: movie-name, theater, number-of-tickets, price, o Inspired by speech act theory (communication as action) request, confirm, inform, thank-you, o Some may take parameters: "Is Kungfu Panda the movie you are look...
[]
GEM-SciDuet-train-12#paper-980#slide-2
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-2
A Multi turn Task oriented Dialogue Architecture
Request(movie; actor=bill murray) Knowledge Base When was it released
Request(movie; actor=bill murray) Knowledge Base When was it released
[]
GEM-SciDuet-train-12#paper-980#slide-3
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-3
A unified view dialogue as optimal decision making
Dialogue as a Markov Decision Process (MDP) Given state , select action according to (hierarchical) policy Receive reward , observe new state Continue the cycle until the episode terminates. Goal of dialogue learning: find optimal to maximize expected rewards Dialogue State (s) Action (a) Reward (r) (Q&A bot over KB, W...
Dialogue as a Markov Decision Process (MDP) Given state , select action according to (hierarchical) policy Receive reward , observe new state Continue the cycle until the episode terminates. Goal of dialogue learning: find optimal to maximize expected rewards Dialogue State (s) Action (a) Reward (r) (Q&A bot over KB, W...
[]
GEM-SciDuet-train-12#paper-980#slide-4
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-4
Task completion dialogue as RL
(utterances in natural language form) o +10 upon successful termination o -10 upon unsuccessful termination o -1 per turn o Pioneered by [Levin+ 00] Other early examples: [Singh+ 02; Pietquin+ 04; Williams&Young 07; etc.]
(utterances in natural language form) o +10 upon successful termination o -10 upon unsuccessful termination o -1 per turn o Pioneered by [Levin+ 00] Other early examples: [Singh+ 02; Pietquin+ 04; Williams&Young 07; etc.]
[]
GEM-SciDuet-train-12#paper-980#slide-5
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-5
RL vs SL supervised learning
Differences from supervised learning Learn by trial-and-error (experimenting) Optimize long-term reward (1 Need temporal credit assignment Similarities to supervised learning input/feature Generalization and representation SL Hierarchical problem solving
Differences from supervised learning Learn by trial-and-error (experimenting) Optimize long-term reward (1 Need temporal credit assignment Similarities to supervised learning input/feature Generalization and representation SL Hierarchical problem solving
[]
GEM-SciDuet-train-12#paper-980#slide-6
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-6
Learning w real users
- Expensive: need large amounts of real experience except for very simple tasks - Risky: bad experiences (during exploration) drive users away
- Expensive: need large amounts of real experience except for very simple tasks - Risky: bad experiences (during exploration) drive users away
[]
GEM-SciDuet-train-12#paper-980#slide-7
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-7
Learning w user simulators
- Inexpensive: generate large amounts of simulated experience for free - Overfitting: discrepancy btw real users and simulators Dialog agent simulated experience
- Inexpensive: generate large amounts of simulated experience for free - Overfitting: discrepancy btw real users and simulators Dialog agent simulated experience
[]
GEM-SciDuet-train-12#paper-980#slide-8
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-8
Dyna Q integrating planning and learning
combining model-free and model-based RL tabular methods and linear function approximation
combining model-free and model-based RL tabular methods and linear function approximation
[]
GEM-SciDuet-train-12#paper-980#slide-9
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-9
Deep Dyna Q DDQ Integrating Planning for Dialogue Policy Learning
Policy as DNN, trained using DQN Apply to dialogue: simulated user as world model Dialogued agent trained using Limited real user experience Large amounts of simulated experience Acting Direct World model Limited real experience is used to improve RL World model (simulated user) Model learning
Policy as DNN, trained using DQN Apply to dialogue: simulated user as world model Dialogued agent trained using Limited real user experience Large amounts of simulated experience Acting Direct World model Limited real experience is used to improve RL World model (simulated user) Model learning
[]
GEM-SciDuet-train-12#paper-980#slide-12
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-12
Dialogue System Evaluation
Metrics: what numbers matter? o Success rate: #Successful_Dialogues / #All_Dialogues o Average turns: average number of turns in a dialogue o User satisfaction o Consistency, diversity, engaging, ... o Latency, backend retrieval cost, Methodology: how to measure those numbers?
Metrics: what numbers matter? o Success rate: #Successful_Dialogues / #All_Dialogues o Average turns: average number of turns in a dialogue o User satisfaction o Consistency, diversity, engaging, ... o Latency, backend retrieval cost, Methodology: how to measure those numbers?
[]
GEM-SciDuet-train-12#paper-980#slide-13
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-13
Evaluation methodology
(lab, Mechanical Turk, ) (optionally with continuing incremental refinement)
(lab, Mechanical Turk, ) (optionally with continuing incremental refinement)
[]
GEM-SciDuet-train-12#paper-980#slide-15
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-15
Agenda based Simulated User
[Schatzmann & Young 09] User state consists of (agenda, goal); goal is fixed throughout dialogue Agenda is maintained (stochastically) by a first-in-last-out stack Implementation of a simplified user simulator: https://github.com/MiuLab/TC-Bot
[Schatzmann & Young 09] User state consists of (agenda, goal); goal is fixed throughout dialogue Agenda is maintained (stochastically) by a first-in-last-out stack Implementation of a simplified user simulator: https://github.com/MiuLab/TC-Bot
[]
GEM-SciDuet-train-12#paper-980#slide-16
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-16
Simulated user evaluation
DQN vs DDQ () : number of planning steps (generating K simulated dialogues per real dialogue)
DQN vs DDQ () : number of planning steps (generating K simulated dialogues per real dialogue)
[]
GEM-SciDuet-train-12#paper-980#slide-17
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-17
Impact of world model quality
pretrained on labeled data, and updated using real dialogue on the fly
pretrained on labeled data, and updated using real dialogue on the fly
[]
GEM-SciDuet-train-12#paper-980#slide-18
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-18
Human in the loop experiments learning dialogue via interacting with real users
DDQ agents significantly outperforms the DQN agent A larger leads to more aggressive planning and better results Pre-training world model with human conversational data improves the learning efficiency and the agents performance
DDQ agents significantly outperforms the DQN agent A larger leads to more aggressive planning and better results Pre-training world model with human conversational data improves the learning efficiency and the agents performance
[]
GEM-SciDuet-train-12#paper-980#slide-19
980
Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning
Training a task-completion dialogue agent via reinforcement learning (RL) is costly because it requires many interactions with real users. One common alternative is to use a user simulator. However, a user simulator usually lacks the language complexity of human interlocutors and the biases in its design may tend to de...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dia...
GEM-SciDuet-train-12#paper-980#slide-19
Conclusion and Future Work
Deep Dyna-Q: integrating planning for dialogue policy learning Make the best use of limited real user experiences Learning when to switch between real and simulated users Exploration: trying actions to improve the world model Exploitation: trying to behave in the optimal way given the current world model
Deep Dyna-Q: integrating planning for dialogue policy learning Make the best use of limited real user experiences Learning when to switch between real and simulated users Exploration: trying actions to improve the world model Exploitation: trying to behave in the optimal way given the current world model
[]
GEM-SciDuet-train-13#paper-982#slide-1
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-1
Motivation
Hierarchical phrase-based model limit Linguistic features (Japanese) subject object verb structure auxiliary words
Hierarchical phrase-based model limit Linguistic features (Japanese) subject object verb structure auxiliary words
[]
GEM-SciDuet-train-13#paper-982#slide-2
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-2
Verb Case Frame
Deep verb case frame between paralleled sentences in two languages Subject Object Time Location Tool Specific to Japanese explicit case frame Agent Time Object Goal Tool Verb Time Agent Tool Verb Object Goal Deep case frame to shallow case frame for Japanese
Deep verb case frame between paralleled sentences in two languages Subject Object Time Location Tool Specific to Japanese explicit case frame Agent Time Object Goal Tool Verb Time Agent Tool Verb Object Goal Deep case frame to shallow case frame for Japanese
[]
GEM-SciDuet-train-13#paper-982#slide-3
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-3
Method
Case Frame Rule extraction Obtain case frame rules from paralleled sentences with word alignments Transform case frame rules into hiero rules. examples of case frame rules (a) the example of phrase rule transformation (b) the example of reordering rule transformation
Case Frame Rule extraction Obtain case frame rules from paralleled sentences with word alignments Transform case frame rules into hiero rules. examples of case frame rules (a) the example of phrase rule transformation (b) the example of reordering rule transformation
[]
GEM-SciDuet-train-13#paper-982#slide-4
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-4
Experiment
CWMT 2011 Japanese-Chinese Corpus (sentence pairs) ASPEC-JC Corpus (sentence pairs) Training data: 680 thousand exp1: Strong hierarchical phrase-based system (baseline) exp2: exp1 with case frame rules exp3: exp1 with manually case frame rules Variables in rule are without distinction during decoding system system CWMT...
CWMT 2011 Japanese-Chinese Corpus (sentence pairs) ASPEC-JC Corpus (sentence pairs) Training data: 680 thousand exp1: Strong hierarchical phrase-based system (baseline) exp2: exp1 with case frame rules exp3: exp1 with manually case frame rules Variables in rule are without distinction during decoding system system CWMT...
[]
GEM-SciDuet-train-13#paper-982#slide-5
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-5
Conclusion
This paper presented an approach to improve HPB model systems by augmenting the SCFG with Japanese CFRs. The CF are used to introduce new linguistically sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions. We obtain significant imp...
This paper presented an approach to improve HPB model systems by augmenting the SCFG with Japanese CFRs. The CF are used to introduce new linguistically sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions. We obtain significant imp...
[]
GEM-SciDuet-train-13#paper-982#slide-6
982
Integrating Case Frame into Japanese to Chinese Hierarchical Phrase-based Translation Model
This paper presents a novel approach to enhance hierarchical phrase-based (HP-B) machine translation systems with case frame (CF).we integrate the Japanese shallow CF into both rule extraction and decoding. All of these rules are then employed to decode new sentences in Japanese with source language case frame. The res...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule i...
GEM-SciDuet-train-13#paper-982#slide-6
Future work
Soft/hard constraints on case frame rule matching Challenge to resolve the problem of tense and aspect etc.
Soft/hard constraints on case frame rule matching Challenge to resolve the problem of tense and aspect etc.
[]
GEM-SciDuet-train-14#paper-986#slide-0
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-0
Corpus highlights
Slides available at http://bit.ly/cl-scisumm|6-slides and will be filed in GitHub. Continuing effort to advance scientific document summarization by encouraging the incorporation of semantic and citation information. Corpus enlarged from 10 (pilot) to 30 CL articles Annotation by 6 paid and trained annotators from U-Hy...
Slides available at http://bit.ly/cl-scisumm|6-slides and will be filed in GitHub. Continuing effort to advance scientific document summarization by encouraging the incorporation of semantic and citation information. Corpus enlarged from 10 (pilot) to 30 CL articles Annotation by 6 paid and trained annotators from U-Hy...
[]
GEM-SciDuet-train-14#paper-986#slide-1
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-1
Oral Sessions
Slides available at http://bit.ly/cl-scisummi|6-slides and will be filed in GitHub. Rais) nai System 8 Top in Task |B, among top performers for Task |A and Task 2 * Remote presentation from China pose | onl) System 6 Among top performers for Task IA
Slides available at http://bit.ly/cl-scisummi|6-slides and will be filed in GitHub. Rais) nai System 8 Top in Task |B, among top performers for Task |A and Task 2 * Remote presentation from China pose | onl) System 6 Among top performers for Task IA
[]
GEM-SciDuet-train-14#paper-986#slide-2
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-2
Evaluation
Still a work in progress: Will present results based on the CEUR paper (old), stacked average ofall runs... ... and contrast with newer (still preliminary) results (new), individual runs separated Task |A Exact sentence ID match EC asi conditional on Task |A Bag of Words (BOVV) overlap between discourse facets BIRNDL 2...
Still a work in progress: Will present results based on the CEUR paper (old), stacked average ofall runs... ... and contrast with newer (still preliminary) results (new), individual runs separated Task |A Exact sentence ID match EC asi conditional on Task |A Bag of Words (BOVV) overlap between discourse facets BIRNDL 2...
[]
GEM-SciDuet-train-14#paper-986#slide-3
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-3
System Results Task 1A and 1B
es) \ HS VY S S 6 Gy a PS) aS as a ne aS ns as A ro a i of of? of? ro a a of? a of oY PY oy of? of a BIRNDL 2016: CL-SciSumm 16 Overview 23 June 2016 7 CEUR version (all system runs averaged)
es) \ HS VY S S 6 Gy a PS) aS as a ne aS ns as A ro a i of of? of? ro a a of? a of oY PY oy of? of a BIRNDL 2016: CL-SciSumm 16 Overview 23 June 2016 7 CEUR version (all system runs averaged)
[]
GEM-SciDuet-train-14#paper-986#slide-4
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-4
Best Performing System Task 1A
System ID Avg Best performing StDev performance Systems
System ID Avg Best performing StDev performance Systems
[]
GEM-SciDuet-train-14#paper-986#slide-5
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-5
Best Performing System Task 1B
System ID Avg StDev performance Best performing
System ID Avg StDev performance Best performing
[]
GEM-SciDuet-train-14#paper-986#slide-6
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-6
Best Performing System Task 2
System ID Approaches Comments Systems
System ID Approaches Comments Systems
[]
GEM-SciDuet-train-14#paper-986#slide-7
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-7
New Results Task 1A
New Results (Task | A) BIRNDL 2016: CL-SciSumm 16 Overview VER TAO) a System ID Approach Task 1a Comments
New Results (Task | A) BIRNDL 2016: CL-SciSumm 16 Overview VER TAO) a System ID Approach Task 1a Comments
[]
GEM-SciDuet-train-14#paper-986#slide-8
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-8
New Results Task 1B
ID Approach Task 1B
ID Approach Task 1B
[]
GEM-SciDuet-train-14#paper-986#slide-9
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-9
New Results Task 2
New Results Task 2 Peele n ulate kod) TINWAaa$oT MSc eRRCAAicne een a Nts tch Aone Le ests ech Atom CERAM Stch Aine Mee a Si-6co ALL m Pac tco Aine - TNL SST ZNOUSOT Bay ieh- e-L0h a ONTLOASS ZTINISE ONTLOASS BBS Ete Ba feXol Bits qduwoowrss CawoowLss Tek oh cuks take) ONILOASS Bere oR Ao) ante sew tt 7TDWISE (aCe) Rte...
New Results Task 2 Peele n ulate kod) TINWAaa$oT MSc eRRCAAicne een a Nts tch Aone Le ests ech Atom CERAM Stch Aine Mee a Si-6co ALL m Pac tco Aine - TNL SST ZNOUSOT Bay ieh- e-L0h a ONTLOASS ZTINISE ONTLOASS BBS Ete Ba feXol Bits qduwoowrss CawoowLss Tek oh cuks take) ONILOASS Bere oR Ao) ante sew tt 7TDWISE (aCe) Rte...
[]
GEM-SciDuet-train-14#paper-986#slide-10
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-10
Supplemental Analyses
We investigated whether high deviations could be because of the topic Topics with both high and low number of citances have mixed results No significant patterns of performance against: Number of citances of the topic set Age of the paper BIRNDL 2016: CL-SciSumm 16 Overview PEM wA0 Ty a)
We investigated whether high deviations could be because of the topic Topics with both high and low number of citances have mixed results No significant patterns of performance against: Number of citances of the topic set Age of the paper BIRNDL 2016: CL-SciSumm 16 Overview PEM wA0 Ty a)
[]
GEM-SciDuet-train-14#paper-986#slide-11
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-11
Limitations
Task |B: limited number of samples for most (e.g.,hypothesis) discourse facets, inconsistent labeling Preprocessing: OCR + Parsing Software: Protege w/ manual alignment and post-processing Scaling the corpus was difficult: key bottleneck in the corpus development The Corpus size, #citing papers
Task |B: limited number of samples for most (e.g.,hypothesis) discourse facets, inconsistent labeling Preprocessing: OCR + Parsing Software: Protege w/ manual alignment and post-processing Scaling the corpus was difficult: key bottleneck in the corpus development The Corpus size, #citing papers
[]
GEM-SciDuet-train-14#paper-986#slide-13
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-13
Conclusions
Slides available at http://bit.ly/cl-scisummi|6-slides and will be filed in GitHub. Successful enlargement of the 2014 pilot task, albeit with some clarification issues We invite teams to examine the detailed results available with the GitHub repo: https://erthub.com/WING-NUS/scisumm-corpus/ Results and finalized analy...
Slides available at http://bit.ly/cl-scisummi|6-slides and will be filed in GitHub. Successful enlargement of the 2014 pilot task, albeit with some clarification issues We invite teams to examine the detailed results available with the GitHub repo: https://erthub.com/WING-NUS/scisumm-corpus/ Results and finalized analy...
[]
GEM-SciDuet-train-14#paper-986#slide-15
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-15
Scientific Document Summarization
wos 1ie-lell Yom I nal eae Surface, lexical, semantic or rhetorical features of the paper Community creates a summary when citing Capture all aspects of a paper
wos 1ie-lell Yom I nal eae Surface, lexical, semantic or rhetorical features of the paper Community creates a summary when citing Capture all aspects of a paper
[]
GEM-SciDuet-train-14#paper-986#slide-16
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-16
Scientific Document Summarization Citation based extractive summaries
Qazvinian,V.,and Radey, D.R.Identifying non-explicit citing sentences for citation-based summarization (ACL, 2010) Abu-|bara, Amjad, and Dragomir Radev. Reference scope identification in citing sentences. (ACL, 2012) Abu-Jbara, Amjad, and Dragomir Radev. Coherent citation- based summarization of scientific papers. (ACL...
Qazvinian,V.,and Radey, D.R.Identifying non-explicit citing sentences for citation-based summarization (ACL, 2010) Abu-|bara, Amjad, and Dragomir Radev. Reference scope identification in citing sentences. (ACL, 2012) Abu-Jbara, Amjad, and Dragomir Radev. Coherent citation- based summarization of scientific papers. (ACL...
[]
GEM-SciDuet-train-14#paper-986#slide-17
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-17
In summary
Community concurs that a citation-based summary of a scientific document Is important. Citing papers cite different aspects of the same reference paper. Assigning facets to these citances may help create
Community concurs that a citation-based summary of a scientific document Is important. Citing papers cite different aspects of the same reference paper. Assigning facets to these citances may help create
[]
GEM-SciDuet-train-14#paper-986#slide-19
986
Overview of the CL-SciSumm 2016 Shared Task
The CL-SciSumm 2016 Shared Task is the first medium-scale shared task on scientific document summarization in the computational linguistics (CL) domain. The task built off of the experience and training data set created in its namesake pilot task, which was conducted in 2014 by the same organizing committee. The track ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }
GEM-SciDuet-train-14#paper-986#slide-19
Annotating the SciSumm corpus
6 annotators selected from a pool of 25 6 hours of training Gold standard annotations for Task |A and IB, per topic or reference paper Community and hand-written summaries for Task 2,
6 annotators selected from a pool of 25 6 hours of training Gold standard annotations for Task |A and IB, per topic or reference paper Community and hand-written summaries for Task 2,
[]
GEM-SciDuet-train-15#paper-991#slide-0
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-0
Introduction
I Bilingual transfer learning is important for overcoming data sparsity in the target language I Bilingual word embeddings eliminate the gap between source and target language vocabulary I Resources required for bilingual methods are often I Texts for embeddings I Source language training samples I We focused on domain...
I Bilingual transfer learning is important for overcoming data sparsity in the target language I Bilingual word embeddings eliminate the gap between source and target language vocabulary I Resources required for bilingual methods are often I Texts for embeddings I Source language training samples I We focused on domain...
[]
GEM-SciDuet-train-15#paper-991#slide-1
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-1
Motivation
I Cross-lingual sentiment analysis of tweets triste sad awful horrible bad malo super super mug jarra rojo hoy red today I Combination of two methods: I Domain adaptation of bilingual word embeddings I Semi-supervised system for exploiting unlabeled data I No additional annotated resource is needed: I Cross-lingual sen...
I Cross-lingual sentiment analysis of tweets triste sad awful horrible bad malo super super mug jarra rojo hoy red today I Combination of two methods: I Domain adaptation of bilingual word embeddings I Semi-supervised system for exploiting unlabeled data I No additional annotated resource is needed: I Cross-lingual sen...
[]
GEM-SciDuet-train-15#paper-991#slide-2
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-2
Word Embedding Adaptation
Source Out-of-domain In-domain W2V MWE Target Out-of-domain In-domain W2V MWE BWE I Goal: domain-specific bilingual word embeddings with general Monolingual word embeddings on concatenated data I Easily accessible general (out-of-domain) data Map monolingual embeddings to a common space using I Small seed lexicon conta...
Source Out-of-domain In-domain W2V MWE Target Out-of-domain In-domain W2V MWE BWE I Goal: domain-specific bilingual word embeddings with general Monolingual word embeddings on concatenated data I Easily accessible general (out-of-domain) data Map monolingual embeddings to a common space using I Small seed lexicon conta...
[]
GEM-SciDuet-train-15#paper-991#slide-3
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-3
Semi Supervised Approach
I Goal: Unlabeled samples for training I Tailored system from computer vision to NLP (Hausser et al., 2017) I Labeled/unlabeled samples in the same class are similar I Sample representation is given by the n 1th layer I Walking cycles: labeled unlabeled labeled I Maximize the number of correct cycles I L Lclassificatio...
I Goal: Unlabeled samples for training I Tailored system from computer vision to NLP (Hausser et al., 2017) I Labeled/unlabeled samples in the same class are similar I Sample representation is given by the n 1th layer I Walking cycles: labeled unlabeled labeled I Maximize the number of correct cycles I L Lclassificatio...
[]
GEM-SciDuet-train-15#paper-991#slide-4
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-4
Cross Lingual Sentiment Analysis of Tweets
I RepLab 2013 sentiment classification (+/0/-) of En/Es tweets I @churcaballero jajaja con lo bien que iba el volvo... I General domain data: 49.2M OpenSubtitles sentences I Twitter specific data: I 22M downloaded tweets I Seed lexicon: frequent English words from BNC (Kilgarriff, 1997) I Labeled data: RepLab En traini...
I RepLab 2013 sentiment classification (+/0/-) of En/Es tweets I @churcaballero jajaja con lo bien que iba el volvo... I General domain data: 49.2M OpenSubtitles sentences I Twitter specific data: I 22M downloaded tweets I Seed lexicon: frequent English words from BNC (Kilgarriff, 1997) I Labeled data: RepLab En traini...
[]
GEM-SciDuet-train-15#paper-991#slide-5
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-5
Medical Bilingual Lexicon Induction
I Mine Dutch translations of English medical words I General domain data: 2M Europarl (v7) sentences I Medical data: 73.7K medical Wikipedia sentences I Medical seed lexicon (Heyman et al., 2017) En word in BNC 5 most similar and 5 random Du pair En word in medical lexicon 3 most similar Du I Classifier based approach ...
I Mine Dutch translations of English medical words I General domain data: 2M Europarl (v7) sentences I Medical data: 73.7K medical Wikipedia sentences I Medical seed lexicon (Heyman et al., 2017) En word in BNC 5 most similar and 5 random Du pair En word in medical lexicon 3 most similar Du I Classifier based approach ...
[]
GEM-SciDuet-train-15#paper-991#slide-6
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-6
Results Sentiment Analysis
labeled data unlabeled data Table 1: Accuracy on cross-lingual sentiment analysis of tweets
labeled data unlabeled data Table 1: Accuracy on cross-lingual sentiment analysis of tweets
[]
GEM-SciDuet-train-15#paper-991#slide-7
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-7
Results Bilingual Lexicon Induction
labeled lexicon unlabeled lexicon medical BNC medical medical medical Table 2: F1 scores of medical bilingual lexicon induction
labeled lexicon unlabeled lexicon medical BNC medical medical medical Table 2: F1 scores of medical bilingual lexicon induction
[]
GEM-SciDuet-train-15#paper-991#slide-8
991
Two Methods for Domain Adaptation of Bilingual Tasks: Delightfully Simple and Broadly Applicable
Bilingual tasks, such as bilingual lexicon induction and cross-lingual classification, are crucial for overcoming data sparsity in the target language. Resources required for such tasks are often out-of-domain, thus domain adaptation is an important problem here. We make two contributions. First, we test a delightfully...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual ...
GEM-SciDuet-train-15#paper-991#slide-8
Conclusions
I Bilingual transfer learning yield poor results when using I We showed that performance can be increased by using only additional unlabeled monolingual data I Delightfully simple approach to adapt embeddings I Broadly applicable method to exploit unlabeled data I Language and task independent approaches
I Bilingual transfer learning yield poor results when using I We showed that performance can be increased by using only additional unlabeled monolingual data I Delightfully simple approach to adapt embeddings I Broadly applicable method to exploit unlabeled data I Language and task independent approaches
[]
GEM-SciDuet-train-16#paper-994#slide-0
994
Simple and Effective Text Simplification Using Semantic and Neural Methods
Sentence splitting is a major simplification operator. Here we present a simple and efficient splitting algorithm based on an automatic semantic parser. After splitting, the text is amenable for further fine-tuned simplification operations. In particular, we show that neural Machine Translation can be effectively used ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3.1", "3.2", "4", "5", "6", "7", "8" ], "paper_header_content": [ "Introduction", "Related Work", "Semantic Representation", "The Semantic Rules", "Neural Component", "Experimental Setup", "Results", "Additio...
GEM-SciDuet-train-16#paper-994#slide-0
Text Simplification
Last year I read the book John authored John wrote a book. I read the book. Original sentence One or several simpler sentences Multiple motivations Preprocessing for Natural Language Processing tasks e.g., machine translation, relation extraction, parsing Reading aids, Language Comprehension e.g., people with aphasia, ...
Last year I read the book John authored John wrote a book. I read the book. Original sentence One or several simpler sentences Multiple motivations Preprocessing for Natural Language Processing tasks e.g., machine translation, relation extraction, parsing Reading aids, Language Comprehension e.g., people with aphasia, ...
[]
GEM-SciDuet-train-16#paper-994#slide-1
994
Simple and Effective Text Simplification Using Semantic and Neural Methods
Sentence splitting is a major simplification operator. Here we present a simple and efficient splitting algorithm based on an automatic semantic parser. After splitting, the text is amenable for further fine-tuned simplification operations. In particular, we show that neural Machine Translation can be effectively used ...
{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, ...
{ "paper_header_number": [ "1", "2", "3.1", "3.2", "4", "5", "6", "7", "8" ], "paper_header_content": [ "Introduction", "Related Work", "Semantic Representation", "The Semantic Rules", "Neural Component", "Experimental Setup", "Results", "Additio...
GEM-SciDuet-train-16#paper-994#slide-1
In this talk
Compares favorably to the state-of-the-art in combined structural and lexical simplification. The first simplification system combining structural transformations, using semantic structures, and neural machine translation. Alleviates the over-conseratism of MT-based systems.
Compares favorably to the state-of-the-art in combined structural and lexical simplification. The first simplification system combining structural transformations, using semantic structures, and neural machine translation. Alleviates the over-conseratism of MT-based systems.
[]