TL;DR: We have used SQuAD datasets and BERT and RoBERTa models designed for English question answering task. We have retrained them to obtain models solving question answering for the Czech language using cross-lingual transfer.
Longer description: Question answering is concerned with natural language processing and information rtetrieval. The main goal is to build a model that can automatically find an answer to a questions posed by humans in given text. There exists several models and datasets for this task in English but there are none for Czech which is more challenging because of its more complicated grammar and richer vocabulary.
This project focuses on building question answerting systems for Czech without requiring any manually annotated Czech training data. It is based on existing datasets and models developed for English question answering task. We automatically translated SQuAD 1.1 and SQuAD 2.0 datasets to Czech to create train and test data, which we release at this URL. We then trained and evaluated BERT and XLM-RoBERTa models based on Tranformer language model architecture. For such training, we used this interface.
This repository contains other necessary scripts for dataset and model preprocessing. You can read more about this project and its results in our article.
- 01_translate_EN-CZ - to translate text in English in inputfile using LINDAT Translator to Czech
- 02_lemmatize_text - to lemmatize original texts and answers using MorphoDita
- 03_drop_not_found_answers - to select answers after lemmatization where the answer and text exactly match regardless of the word order
- 04_show_difference - to show difference between answers in two files
- 05_create_html_visualization - to visualize questions and answers in the text
- 06_translate_dev_CZ-EN - to translate text from Czech to English using LINDAT Translator
- 07_translate_predictions_EN-CZ - to translate predictions from English to Czech using LINDAT Translator
- 08_lemmatize_predictions - to lemmatize predicted answers using MorphoDita
- 09_lemmatize_dev - to lemmatize development dataset using MorphoDita
- czech-morfflex-pdt-161115.tagger - model for lemmatization
- czech-morfflex-pdt-161115-derinet20.tagger - model for lemmatization
- translated_ansewers_visualization - to visualize translated answers
- lemmatizer - to lemmatize general text
- README - documentation file
- vizualize_data_sizes_after_translation - to visualize the dataset sizes after translation using LINDAT Translator
- vizualize_results_all_models - to evaluate models and create graphs with results
The translated SQuAD1.1. and SQuAD 2.0 datasets are available here.