denis-berezutskiy-lad's picture
Initial version of README
354d75f
metadata
license: apache-2.0

About

This is a dataset for training Russian punctuators/capitalizers via NeMo scripts (https://github.com/NVIDIA/NeMo)

A BERT model already fine-tuned on this dataset can be found here: https://huggingface.co/denis-berezutskiy-lad/lad_transcription_bert_ru_punctuator

Scripts for collecting/updating such a dataset, as well as training/using the model are located here: https://github.com/denis-berezutskiy-lad/transcription-bert-ru-punctuator-scripts/tree/main

The idea behind the project is to use large continous professional transcriptions for training rather than relying on short low-quality samples consisting of 1-2 sentences (which is typical for the most popular datasets in Russian). Our experiments show significant improvements comparing to BERTs trained on the standard Ru datasets (social comments, omnia russica etc.). That's why we mainly use transcriptions published by Russian legislatures (Gosduma, Mosgorduma) with some addition of film subtitles from OpenSubtitles project.

The dataset is in .csv format, but can be easily converted to the NeMo format (text.txt and labels.txt) - see the custom scripts above.

About 1.2 GB of data is from Gosduma, ~300 MB from Mosgorduma and ~300 MB from a snapshot of Russian OpenSubtitles, taken from the Taiga project (https://tatianashavrina.github.io/taiga_site/downloads.html ).

The rows are ordered randomly by the source document ("source_entity" column), however within a document the order is preserved like it was in the original text. So if you need to regroup the texts (for example, make longer or shorter sequences), you may concatenate rows grouped by the source entity and then split them as you want.

Supported labels

Please note that some new labels are not supported by NeMo scripts out of the box (-, —, T), so we need to add special handling for them. See the inference notebook for details.

Punctuation

O,.?!:;…⁈-—

Capitalization

OUT

(T means abbreviation ("total" uppercase))