Scholarlink

Automatic semantic shift detection can be used in a number of applied problems in linguistics, from compiling explanatory dictionaries to analyzing historical documents. In this paper, we consider the related problem of word-sense disambiguation (WSD) and the feasibility of adapting algorithms based on explanatory dictionaries to detect out-of-dictionary meanings. We selected 50 reference words, for which we collected a dataset of context-gloss pairs based on the National Media subcorpus of the Russian National Corpus, news feed of the Vkontakte social network, and the Big Explanatory Dictionary of the Russian Language. We adapted four algorithms originally developed for the WSD problem. The best quality (F1-score = 0.96) was achieved through the GlossBERT algorithm. We also assessed the generalization ability of this algorithm applying a train-test split by reference words. In this case, F1-score dropped significantly to a value of 0.71.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Out-of-Dictionary Meanings Detecting Using Word-Sense Disambiguation Algorithms

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Darya Borodina,
  • Dmitry Morozov

摘要

Automatic semantic shift detection can be used in a number of applied problems in linguistics, from compiling explanatory dictionaries to analyzing historical documents. In this paper, we consider the related problem of word-sense disambiguation (WSD) and the feasibility of adapting algorithms based on explanatory dictionaries to detect out-of-dictionary meanings. We selected 50 reference words, for which we collected a dataset of context-gloss pairs based on the National Media subcorpus of the Russian National Corpus, news feed of the Vkontakte social network, and the Big Explanatory Dictionary of the Russian Language. We adapted four algorithms originally developed for the WSD problem. The best quality (F1-score = 0.96) was achieved through the GlossBERT algorithm. We also assessed the generalization ability of this algorithm applying a train-test split by reference words. In this case, F1-score dropped significantly to a value of 0.71.