Exploring the Impact of Error Type Features Integration on Transformer-based Myanmar Spelling Correction

Spelling correction is a fundamental aspect of natural language processing systems, essential for ensuring the accuracy and fluency of text-based applications. In the context of neural machine translation based spelling correction, spelling errors pose significant challenges, particularly in low-resource language like Myanmar language where training data is limited. In this paper, the Myanmar spelling error-correction parallel corpus was developed through the collection of text from social media platforms. By learning the error patterns in social media text data, we categorized the Myanmar spelling mistakes into ten categories. We developed Myanmar spelling correction model using Transformer architecture and also investigated the potential improvement of our Transformer-based spelling correction by incorporating error type features as additional linguistic information. Evaluating the Transformer model with manually annotated error type labels in our developed corpus significantly enhances the F1-score from 75.91% to 86.71%, increases the GLEU score from 0.9648 to 0.9767, and decreases the word error rate from 3.3% to 2.0%. Additionally, we conducted a preliminary study on integrating our proposed correction model with an automatic error type classification model, employing a bidirectional Long Short-Term Memory. The findings demonstrate that integrating error type information into the Transformer architecture enhances the effectiveness of Myanmar spelling correction.