tommasobonomo/sem_augmented_fever_nli

Semantically-augmented FEVER for NLI

This dataset is a random downsample of the FEVER dataset adapted for NLI. We downsampled the training and development set to 25% of the original, and recovered labels for the development set from the original FEVER dataset.

The dataset is also augmented with semantic annotations such as Word Sense Disambiguation (WSD) and Semantic Role Labeling (SRL) information. We annotated the whole downsampled dataset (both premise and hypothesis) with AMuSE-WSD (Orlando et al., EMNLP 2021) and InVeRo (Conia et al., EMNLP 2020).

Dataset Creation

The idea was to curate a version of the FEVER dataset adapted to the NLI task for Homework 2 of the Multilingual Natural Language Processing 2024 course at Sapienza University of Rome.

We sourced the data following the instructions in this repo, modifying the labels to the following schema:

{
  "id": ..., # the FEVER dataset ID
  "premise": ..., # the context in FEVER
  "hypothesis": ..., # the query in FEVER
  "label": ..., # mapped version of FEVER, where 'supports' -> 'entailment', 'refutes' -> 'contradiction' and 'not enough info' -> 'neutral'
}

We filtered out any sample with an empty premise (a minority of the data) and downsampled both train_fitems.jsonl and dev_fitems.jsonl to 25% of their total size. We then recovered the labels for the development set from the labelled_dev split in the original FEVER and split the development set into the final dev.jsonl and test.jsonl.

Finally, we ran both AMuSE-WSD and InVeRo to augment our samples with WSD and SRL annotations, respectively.