Nengfei Luo, Shanli Ye

DOI: 10.26480/ess.01.2025.01.06

ABSTRACT
Chinese named entity recognition faces unique challenges due to the lack of natural separators such as spaces, making entity boundary determination more ambiguous compared to English. Addressing the issue of noise introduced by lexicon information integration, which often affects model performance, this study proposes an effective multi-task learning approach for lexical enhancement through word selection. Task 1 involves learning a scoring model to select the most useful K words from the matched lexicon entries, while Task 2 integrates character-level and word-level features into a BiLSTM model for sequence labeling. Both tasks are jointly trained using a shared encoder. Experiments conducted on public datasets Weibo, Resume, and MSRA achieved F1-scores of 73.26%, 96.51%, and 95.77%, respectively, demonstrating improved performance over several advanced baseline models. This approach effectively incorporates lexicon information while mitigating the impact of noisy words, thereby enhancing the accuracy of named entity recognition.

KEYWORDS
Named entity recognition,lexical enhancement,scoring model, multi-task learning