Improved grammatical error correction by ranking elementary edits

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11416 - 11429 December 7-11, 2022 ©2022 Association for Computational Linguistics Improved grammatical error correction by ranking elementary edits Alexey Sorokin Yandex Moscow State University alexey dot sorokin at list dot ru Abstract We offer a two-stage reranking method for grammatical error correction: the first model serves as edit generator, while the second clas- sifies the proposed edits as correct or false. We show how to use both encoder-decoder and se- quence labeling models for the first step of our pipeline. We achieve state-of-the-art quality on BEA 2019 English dataset even using weak BERT-GEC edit generator. Combining our roberta-base scorer with state-of-the-art GEC- ToR edit generator, we surpass GECToR by 2 − 3%. With a larger model we establish a new SOTA on BEA development and test sets. Our model also sets a new SOTA on Russian, despite using smaller models and less data than the previous approaches. 1 Introduction Grammatical error correction (GEC) is the task of converting the source text to its clean version with no orthographic, punctuation, lexical or other er- rors. It becomes increasingly popular during last years due to its applications such as Second Lan- guage Learning. However, even for English with its numerous resources and wide research commu- nity, modern models are far from being perfect. In particular, the recall of the state-of-the-art GEC- ToR model (Omelianchuk et al., 2020) on stan- dard BEA2019 development set is lower than 40%. While GECToR uses sequence labeling approach with linguistically motivated label inventory, most works for other languages, such as (Náplava and Straka, 2019) and others, simply mimic machine translation methods, training a Transformer model (Vaswani et al., 2017) on the pairs of source and corrected sentences. This approach suffers from left-to-right decoding: the model can make a wrong decision not observing the future context. This problem may be mitigated using reranking: the ranker observes entire corrected sequences and thus may utilize richer context. It also helps to discriminate between several possible edits with similar basic model probability . Due to these rea- sons, it was heavily used in machine translation both in statistical (Och et al., 2004) and neural (Yee et al., 2019) era. In contrast to machine translation, sequence edit- ing in GEC usually does not require complete rewriting and can be decomposed to elementary edits such as modifying a single word or a consec- utive group of words. In this paper we propose to score elementary edits produced by the basic model and classify them as positive or negative on the second stage of the pipeline. Than the calcu- lated probabilities can be either used directly or combined with the scores from the first stage. We show that our scoring model achieves state- of-the-art performance on BEA2019 dataset even with rather weak first stage model. Its combina- tion with GECToR edit generator outperforms the models of the same size by about 2 F0.5 points. The large version of our model …

Content truncated to 500 words