Non-randomness in Morphological Diversity: A Computational Approach Based on Multilingual Corpora

Abstract

The project applies information theory, statistical modelling and machine learning to the study of language adaptation using linguistic data extracted from multilingual corpora. In addition to the theoretical findings, the project will provide a data set consisting of text samples of 100 languages facilitating future use of corpus-based computational methods in scientific approaches to linguistic diversity and change.

Project members:  Olga Sozinova (PhD student), Ximena Gutierrez-Vasques (PostDoc), Christian Bentz (PostDoc, external collaborator), Steven Moran (PostDoc, external collaborator) and Tanja Samardžić (PI).

Funding:  SNF grant #176305 2018—2022. 

Computational methods to describe languages

Publication in the EU Research magazine about our project, aiming at a wide audience. Spring 2022.

Link to the full issue

EU Research publication

 

TeDDi Sample: Text Data Diversity Sample for Language Comparison and Multilingual NLP

Steven Moran, Christian Bentz, Ximena Gutierrez-Vasques, Olga Sozinova and Tanja Samardzic. 2022. "TeDDi Sample: Text Data Diversity Sample for Language Comparison and Multilingual NLP”. In Proceedings of The International Conference on Language Resources and Evaluation (LREC), Marseille, France. 20—25 June 2022.

GitHub repository Paper (to appear) Video (to appear)

 

Collecting the TeDDi Sample

The turning point of BPE merges

Ximena Gutierrez-Vasques, Christian Bentz, Olga Sozinova and Tanja Samardzic. 2021. "From characters to words: the turning point of BPE merges”. European Chapter of the Association for Computational Linguistics, Long Papers.

Paper GitHub repository Twitter thread  Video

Interpretability for morphological inflection

Tatyana Ruzsics, Olga Sozinova, Ximena Gutierrez-Vasques and Tanja Samardzic. 2021. "Interpretability for morphological inflection: from character-level predictions to subword-level rules”. European Chapter of the Association for Computational Linguistics, Long Papers.

Paper GitHub repository  Video

100-language sample

Data for the analysis consists of texts in 100 languages, which will be published as a multilingual corpus. The chosen 100-language sample is proposed by WALS. The text collection for this sample of languages will be our original contribution.

Current collection

Current number of tokens collected per genre: fiction, non-fiction, conversation, professional, technical, grammar examples.

Full page view