CorefPM: Coreference Resolution as Linguistic Prompt Prediction and Machine Reading Comprehension
Jianxuan Zhao, Huazhen Wang, Yifei Zhao · 2022
Coreference resolution is a key challenge in natural language processing and is considered as a good test of machine intelligence. Recent coreference resolution prefers to supervised methods due to its superior performance. But supervised methods suffer from data scarcity and limited generalizability. In this paper, we present CorefPM, an unsupervised approach for the coreference resolution task. We formulate the task as following steps: (1) We generate the candidate mention set by linguistic knowledge and score them by prompt prediction method.(2) We obtain another mention with its confidence score by machine reading comprehension(MRC) model.(3) We output final resolution result by comparing candidate mention and MRC mention. Experiments demonstrate that our approach significantly outperforms the baseline of few-shot and zero-shot in ZeroCLUE, although performs worse on CLUEWSC 2020 dataset than supervised learning baselines. Our experiment results suggested that linguistic knowledge is valuable and can improve the result of coreference resolution task.