结合注意力机制的多策略汉语语义角色标注

(1.西北民族大学中国民族语言文字信息技术教育部重点实验室,甘肃 兰州 730030; 2.西北民族大学甘肃省民族语言智能处理重点实验室,甘肃 兰州 730030)

汉语语义角色标注; 双向长短时记忆; 条件随机场; 注意力机制; 依存句法分析; 短语结构句法分析

Multi-strategy Chinese semantic role labeling combined with attention mechanism
ZHU Ao1,WAN Fucheng1,2*,MA Ning2,CHE Guoyi1

(1.Key Laboratory of China's Ethnic Languages and Information Technology of Ministry of Education,Northwest Minzu University,Lanzhou 730030,China; 2.Key Laboratory of China's Ethnic Languages and Intelligent Processing of Gansu Province,Northwest Minzu Un

Chinese semantic role labeling; BiLSTM; CRF; attention mechanism; dependency syntactic parsing; phrase structure parsing

DOI: 10.6043/j.issn.0438-0479.202012005

备注

语义角色标注旨在标注出句子中所有与谓语相关的语义成分,是进行语义分析的基础和关键技术.使用传统的机器学习方法进行语义角色标注,需要人工设定特征,特征稀疏且工作繁琐沉重,同时传统方法对句法解析精度有较高要求,所以语义角色标注发展缓慢.针对上述情况,采取基于双向长短时记忆(BiLSTM)网络-注意力机制(attention)-条件随机场(CRF)模型进行汉语语义角色标注,同时尝试针对性引入其他资源优化模型性能.在训练阶段,将词性、依存句法特征以及短语结构句法特征组成的多线索特征组共同送入模型.经过多组对照实验论证,相比于BiLSTM-CRF模型,融合注意力机制的模型性能显著提升,并且引入的多线索特征组可以进一步提升模型性能.
Semantic role labeling aims to label all semantic components related to predicates in a sentence,and it constitutes the basis and the key technology for semantic analyses.Semantic role labeling based on traditional machine learning methods requires manual feature setting.Traditional methods have endured problems of sparse features and cumbersome work.At the same time,traditional methods must meet higher requirements on the accuracy of syntactic analyses,rendering the development of semantic role labeling slow.Due to the existence of the situation mentioned above,this paper is based on the BiLSTM-attention-CRF model for Chinese semantic role labeling,and attempts to introduce other targeted resources to optimize the performance of the model.During training,a multi-clue feature group consisting of parts of speech,dependent syntactic parsing features,and phrase structure parsing features are fed into the model.After multiple groups of controlled experiments,compared with the BiLSTM-CRF model,the performance of the model fused with the attention mechanism is significantly improved.Finally,the combined multi-cue feature group can further improve the model performance.