Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review
Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review
}
TY - GEN
T1 - The Impact of Cross-Lingual Adjustment of Contextual Word Representations on Zero-Shot Transfer
T2 - book chapter
AU - Efimov, Pavel
AU - Boytsov, Leonid
AU - Arslanova, Elena
AU - Braslavski, Pavel
N1 - This research was supported in part through computational resources of HPC facilities at HSE University [27]. PE is grateful to Yandex Cloud for their grant toward computing resources of Yandex DataSphere. PB acknowledges support by the Russian Science Foundation, grant No 20-11-20166.
PY - 2023/3/16
Y1 - 2023/3/16
N2 - Large multilingual language models such as mBERT or XLM-R enable zero-shot cross-lingual transfer in various IR and NLP tasks. Cao et al. [8] proposed a data- and compute-efficient method for cross-lingual adjustment of mBERT that uses a small parallel corpus to make embeddings of related words across languages similar to each other. They showed it to be effective in NLI for five European languages. In contrast we experiment with a topologically diverse set of languages (Spanish, Russian, Vietnamese, and Hindi) and extend their original implementations to new tasks (XSR, NER, and QA) and an additional training regime (continual learning). Our study reproduced gains in NLI for four languages, showed improved NER, XSR, and cross-lingual QA results in three languages (though some cross-lingual QA gains were not statistically significant), while mono-lingual QA performance never improved and sometimes degraded. Analysis of distances between contextualized embeddings of related and unrelated words (across languages) showed that fine-tuning leads to “forgetting” some of the cross-lingual alignment information. Based on this observation, we further improved NLI performance using continual learning. Our software is publicly available https://github.com/pefimov/cross-lingual-adjustment.
AB - Large multilingual language models such as mBERT or XLM-R enable zero-shot cross-lingual transfer in various IR and NLP tasks. Cao et al. [8] proposed a data- and compute-efficient method for cross-lingual adjustment of mBERT that uses a small parallel corpus to make embeddings of related words across languages similar to each other. They showed it to be effective in NLI for five European languages. In contrast we experiment with a topologically diverse set of languages (Spanish, Russian, Vietnamese, and Hindi) and extend their original implementations to new tasks (XSR, NER, and QA) and an additional training regime (continual learning). Our study reproduced gains in NLI for four languages, showed improved NER, XSR, and cross-lingual QA results in three languages (though some cross-lingual QA gains were not statistically significant), while mono-lingual QA performance never improved and sometimes degraded. Analysis of distances between contextualized embeddings of related and unrelated words (across languages) showed that fine-tuning leads to “forgetting” some of the cross-lingual alignment information. Based on this observation, we further improved NLI performance using continual learning. Our software is publicly available https://github.com/pefimov/cross-lingual-adjustment.
UR - http://www.scopus.com/inward/record.url?partnerID=8YFLogxK&scp=85151051828
UR - https://gateway.webofknowledge.com/gateway/Gateway.cgi?GWVersion=2&SrcAuth=tsmetrics&SrcApp=tsm_test&DestApp=WOS_CPL&DestLinkType=FullRecord&KeyUT=000995495200004
U2 - 10.1007/978-3-031-28241-6_4
DO - 10.1007/978-3-031-28241-6_4
M3 - Conference contribution
SN - 978-3-031-28237-9
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 51
EP - 67
BT - Advances in Information Retrieval: 45th European Conference on Information Retrieval
A2 - Kamps, Jaap
A2 - Goeuriot, Lorraine
PB - Springer Cham
ER -
ID: 37140299