Abstract
AbstractSpelling error correction is a task in which errors in a natural language sentence can be detected and corrected. In this paper, we consider Chinese spelling error correction (CSC) for generality. A previous state-of-the-art method for this task connects a detection network with a correction network based on BERT by soft masking. This method does solve the problem that BERT has the insufficient capability to detect the position of errors. However, we find that it still lacks sufficient inference ability and world knowledge by analyzing its results. To solve this issue, we propose a novel correction approach based on knowledge graphs (KGs), which queries triples from KGs and injects them into the sentences as domain knowledge. Moreover, we leverage MLM as correction to improve the inference ability of BERT and adopt a denoising filter to increase the accuracy of results. Experimental results on the SIGHAN dataset verify that the performance of our approach is better than state-of-the-art methods.KeywordsSpelling error correctionKnowledge graphBERT
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have