A Study on Context-Matching-Based Joint Training for Chinese Coreference Resolution

To address issues such as ineffective control of the number of mentions and neglecting contextual relationships during mention clustering in end-to-end Chinese coreference resolution models, a context-matching-based joint training Chinese coreference resolution model is proposed. The model utilizes RoBERTa(wwm)-large combined with BiLSTM to encode Chinese text, then clusters word embeddings. It uses the results of the word clustering to recognize span mentions, reducing the number of span mentions that need to be processed. Finally, it reclusters the recognized span mentions. During training, the clustering and span mention recognition stages are jointly trained. Furthermore, during clustering, the matching degree of candidate antecedents and mentions within their respective contexts is considered, forming a clustering scoring system that integrates contextual matching.Experiments show that the model achieves an F1 score of 71.11% on the OntoNotes-5.0 Chinese dataset for the CoNLL metric, an improvement of 0.95% compared to the baseline model. On the self-constructed "Tibet News Traffic" dataset, the F1 score reaches 73.62%, an increase of 2.58% compared to the baseline model. These results indicate that the proposed model can effectively enhance the performance of coreference resolution models and demonstrates good generalization capability.