grammatek/multilingual-pl-bert-is-updated

Introduction

This dataset, derived from the Icelandic Gigaword Corpus, is designed as a more comprehensive alternative to the existing dataset found at https://huggingface.co/datasets/styletts2-community/multilingual-pl-bert/tree/main/is. The original dataset, derived from just 52MB of raw text from the Icelandic Wikipedia, was processed using the espeak-ng backend for normalization and phonemization. However, the Icelandic module of espeak-ng, which has not been updated for over a decade, employs an outdated IPA dialect and a simplistic approach to stress marking. Additionally, the limited phonemization capabilities of the module independently contribute to inaccuracies in the phonetic transcriptions.

Significant advancements in the normalization and G2P (Grapheme-to-Phoneme) conversion of Icelandic have been made through the Icelandic Language Technology program. More information about this program can be found here. The tools developed in this program have been extensively used to enhance the quality of this dataset.

Dataset

This dataset surpasses its predecessor considerably in size, incorporating not only text from the relatively small Icelandic Wikipedia but also from the extensive Icelandic Gigaword corpus. Specifically, we have enriched the Wikipedia text with material from the News1 corpus. To adhere to the maximum size limit of 512 MB for the raw text, we combined the complete Wikipedia text with randomly shuffled documents from the News1 corpus until reaching the size cap.

In total, the dataset contains 400.676 rows, each corresponding to its associated document in the IGC corpus' XML file.

Cleaning

Prior to processing with the Bert tokenizer, the dataset underwent cleaning, deduplication, and language detection to filter out most non-Icelandic text. Documents containing fewer than 10 words were also removed. This preprocessing resulted in the elimination of 8,146 documents from the initial 55,475 in the Wikipedia corpus (approximately 14.7%) and 28,869 from 1,545,671 in the News1 corpus (about 1.9%). The notably higher reduction in the Wikipedia corpus primarily arose from the minimum word count criterion. However, this did not significantly diminish the total volume of text, which only saw a modest decrease from 52.3MB to 49.68MB, a reduction of around 5%.

Normalization

For normalization, we adapted the Regina Normalizer, which employs a BI-LSTM Part-of-Speech (PoS) tagger. Although this makes the process somewhat time-consuming, the adaptions were necessary to handle a variety of edge cases in the diverse and sometimes unclean text within the IGC. The processing of approximately 2.5 GB of raw text took about one day, utilizing 50 CPU cores.

Phonemization

Phonemization was conducted using IceG2P, which is also based on a BI-LSTM model. We made adaptations to ensure the IPA phoneset output aligns with the overall phoneset used in other PL-Bert datasets. Initially, we created and refined a new vocabulary from both the normalized Wikipedia and News1 corpora. Following this, the BI-LSTM model was employed to generate phonetic transcriptions for the dictionary. We also enhanced stress labeling and incorporated secondary stresses after conducting compound analysis.

A significant byproduct of this effort is a considerably improved G2P dictionary with more than 2.1 million transcriptions, which we plan to integrate into the G2P module and various other open-source projects involving Icelandic G2P.

Ultimately, to ensure textual coherence, all paragraphs with incorrect Grapheme-to-Phoneme (G2P) transcriptions were excluded from the dataset.

License

The dataset is distributed under the same CC-by-4.0 license as the original source material from which the data was derived.