Thai Grammatical Error Correction for Deaf Students

Abstract

Deaf students encounter challenges in written communication due to errors such as insertion, deletion, disorder, misusage, and misspellings. Grammatical error correction (GEC) technology can help mitigate these issues. However, existing GEC models are primarily trained on online resources from second-language hearing learners. In contrast, sentences written by deaf students suffer from a variety of errors not typically found elsewhere.To address this issue, we create the Thai Deaf Corpus (TDC), focusing on identifying and analyzing errors among deaf students in grades 7-12 across four deaf schools. Additionally, we introduce a two-stage system for the Thai-GEC model, automatically detecting and correcting incorrect words in ungrammatical sentences written by deaf students.In our experiment, we compare the performance of the recurrent neural networks (RNN) detection model with and without feature embeddings from different sources—TDC and News corpus; moreover, we compare three correction models: WangchanBERTa and Seq2Seq models with and without pretraining. Our analysis of the TDC shows that deaf students use simpler words and shorter sentences compared to hearing peers, who often use complex vocabulary and lengthy sentences. The RNN detection model with feature embeddings learned from only TDC can outperform the others. The Seq2Seq correction model with pretraining outperforms others as it learns from formal and grammatical sentences in the News corpus. The dataset is made available at https://github.com/Supachan/ThaiDeafCorpus.git