基于动态语义注意力的指代消解方法  

Coreference Resolution Based on Dynamic Semantic Attention

作　　者：邓思艺 乐小虬 Deng Siyi;Le Xiaoqiu(National Science Library,Chinese Academy of Sciences,Beijing 100190,China;Department of Library,Information and Archives Management,School of Economics and Management,University of Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院文献情报中心,北京100190 [2]中国科学院大学经济与管理学院图书情报与档案管理系,北京100190

出　　处：《数据分析与知识发现》2020年第5期46-53,共8页Data Analysis and Knowledge Discovery

摘　　要：【目的】针对先行表述复杂、指代词语义不明的问题,探索更有效的指代消解方法。【方法】采用端到端的框架,使用打分排序法识别指代关系。先对文本段中的连续词序列进行"提及"打分,判断是否为"提及";然后利用筛选出的候选"提及"对指代关系打分。其中词序列建模采用动态语义注意力机制,引入更匹配当前指代关系的外部词语义,并使用内部注意力编码,突出先行表述中与指代词关联的部分;综合两部分打分排序得到识别结果。【结果】在基于OntoNotes5.0语料库的CoNLL-2012共享任务英语数据上进行实验,同参数情况下,准确率、召回率、F1值分别比基准模型提高2.02%、0.42%、1.14%。【局限】外部语义表征的来源语料不够丰富,有待补充。训练语料皆为新闻、脱口秀或者网络日志等通用文本,可考虑加入科技文献语料,构造更为丰富的指代情境,并评估模型在各种指代情境下的表现。【结论】动态语义注意力模块可在构建词序列表示时注入更有利于当前指代关系识别的语义特征,动态的、有选择性的外部语义注入更有利于指代关系的识别。[Objective] This paper tries to more effectively identify the coreference, aiming to address the issues of ambiguous anaphor meaning and complex antecedent structure. [Methods] We established an end-to-end framework and used score ranking to identify the coreference relationships. Firstly, we calculated scores of all spans to retrieve the"mentions". Then, we used scores of the candidate mention pairs to determine coreference relationship. We also built span representation with external multiple semantic representations. Finally, we combined scores of the two parts to generate the final list. [Results] We examined our model with the OntoNotes benchmark datasets. The precision, recall and F1 values of our model were 2. 02%, 0. 42% and 1. 14% higher than those of the SOTA model. [Limitations] The training data sets only collected news, talk shows, or weblogs.More sci-tech literature is needed to further improve the model’s performance. [Conclusions] The proposed model could more effectively identify coreferences.

关 键 词：指代消解 动态语义注意力 打分排序模型 深度学习 

分 类 号：G35[文化科学—情报学]