Developing a Thai Grammatical Error Correction Tool for Deaf Students

Deaf students face challenges in written communication due to errors such as insertion, deletion, word order issues, 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 learners who are hearing. In contrast, sentences written by deaf students exhibit a variety of errors not typically found elsewhere. To address this, we aimed to create the Thai Deaf Corpus (TDC) from deaf students in grades 7-12 across four deaf schools. Our analysis of the TDC revealed that deaf students wrote short sentences, averaging six words each, used 4,585 unique words, and predominantly produced ungrammatical sentences. In addition, we introduce a two-stage system (Thai-GEC model) to automatically detect and correct incorrect words in ungrammatical sentences. In our experiments, we compared different detection and correction models on the dataset. As a result, off-the-shelf models perform poorly compared to models specifically created using our corpus, showing the usefulness of our dataset. The TDC is available at https://github.com/Supachan/ThaiDeafCorpus.git.