Spelling Correction for Russian: A Comparative Study of Datasets and Methods

Proceedings of Recent Advances in Natural Language Processing, pages 1206–1216 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_136 1206 Spelling Correction for Russian: A Comparative Study of Datasets and Methods Alla Rozovskaya Department of Computer Science Queens College, CUNY arozovskaya@qc.cuny.edu Abstract We develop a minimally-supervised model for spelling correction and evaluate its perfor- mance on three datasets annotated for spelling errors in Russian. The first corpus is a dataset of Russian social media data that was recently used in a shared task on Russian spelling cor- rection. The other two corpora contain texts produced by learners of Russian as a foreign language. Evaluating on three diverse datasets allows for a cross-corpus comparison. We compare the performance of the minimally- supervised model to two baseline models that do not use context for candidate re-ranking, as well as to a character-level statistical ma- chine translation system with context-based re-ranking. We show that the minimally- supervised model outperforms all of the other models. We also present an analysis of the spelling errors and discuss the difficulty of the task compared to the spelling correction prob- lem in English. 1 Introduction The spelling correction task has been a fundamen- tal Natural Language Processing (NLP) problem ever since the origins of the field and has en- joyed a lot of attention in the NLP research. It is not surprising, since correcting spelling mis- takes is of practical relevance for various higher- level NLP tasks and downstream applications dealing with noisy data, such as named entity recognition, dependency parsing, information re- trieval, topic modeling, machine translation, es- say scoring, speech recognition, automatic text correction (van Strien et al., 2020). Running a spellchecker is now a common pre-processing step performed in essay scoring (Flor, 2012a), grammatical error correction (Chollampatt and Ng, 2018; Rozovskaya and Roth, 2016; Grund- kiewicz and Junczys-Dowmunt, 2018) and numer- ous other applications. Nevertheless, even for En- glish, performance of spellchecking tools is not as good as one would expect, especially in noisy do- mains (Flor et al., 2019). Kantor et al. (2019) eval- uate three publicly available spellcheckers on En- glish learner data and find that the highest recall achieved is that of 69%, and the best precision is 57%, which indicates that the task is far from be- ing solved. Further, their simple in-house imple- mentation outperforms by a large margin all of the common publicly available spellcheckers. One reason for the slow progress on the task might be the lack of common benchmark datasets. As a result, proposed methods are being evaluated either on isolated spelling errors extracted from a corpus without context1 or on artificially cre- ated datasets. Recently, Flor et al. (2019) re- leased a dataset annotated for spelling errors in English learner essays and provided an evaluation of a minimal supervision system that combines features based on the misspelled word itself and the context in which it appears. They report strong performance on that corpus, as well as competitive results on a dataset from the medical domain. We address the problem of …

Content truncated to 500 words