语义角色标注增强的隐式篇章关系识别  

Implicit discourse relation recognition as SRL-enhanced

作　　者：吴一珩 李军辉[1] 朱慕华 WU Yiheng;LI Junhui;ZHU Muhua(School of Computer Science and Technology,Soochow University,Suzhou 215006,China;Beijing Sankuai Online Technology Co.Ltd.,Beijing 100089,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]北京三快在线科技有限公司,北京100089

出　　处：《厦门大学学报（自然科学版）》2024年第6期1005-1015,共11页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(61876120)。

摘　　要：[目的]为了改善因将隐式篇章关系识别(implicit discourse relation recognition,IDRR)描述为文本分类任务,提出了一种新颖的IDRR方法.[方法]将IDRR视为文本生成任务并直接生成篇章单元对的连接词,随后生成的连接词被准确无歧义地映射到唯一篇章关系.具体地,首先设计了可以将无歧义性连接词转化为对应篇章关系的连接词-关系映射表;然后介绍了两种不同的连接词替换策略用于替换训练样例中的歧义性连接词;最后,将IDRR视作一个序列到序列的任务,其中目标端序列由基于语义角色标注增强的篇章单元对和两者之间的连接词组成.[结果]基于英语PDTB和中文CDTB的实验结果表明本文提出的方法达到了最先进的性能.[结论]本文方法创新性地将IDRR视为文本生成任务,并通过序列到序列模型显著提升了隐式篇章关系识别的效果,为隐含信息的精确捕捉提供了新的解决方案.[Objective] Implicit discourse relation recognition(IDRR) has traditionally been formalized as a classification task.In this study,we propose a novel approach that redefines IDRR as a text generation process,thus enabling direct generation of connective words between discourse units.This redefinition allows unambiguous mapping of connectives to discourse relations,aiming to improve precision in IDRR tasks.[Methods] Our approach treats IDRR as a sequence-to-sequence task.A connective-to-relation table is designed to map unambiguous connectives to specific discourse relations.Next,two substitution strategies are developed to replace ambiguous connectives in training instances.Furthermore,discourse units(DUs) are enriched with semantic role labels(SRL),providing the additional context.Finally,the model generates connectives based on these enhanced DUs.[Results] Experimental results on the English PDTB and Chinese CDTB datasets validate the effectiveness of this approach.The proposed method achieves state-of-the-art performance,significantly surpassing previous models in recognizing implicit discourse relations.Effectively,the connective generation resolves ambiguity and directly maps generated connectives to discourse relations.The substitution strategies and the connective-to-relation table enhance accuracy by ensuring unambiguous relational mapping.Overall,this approach demonstrates improved performance in capturing subtle semantic nuances within discourse.[Conclusion] Redefining IDRR as a generation task has shown substantial advantages in accurately capturing implicit discourse relations.By incorporating semantic role labeling and connective mapping strategies,this approach aligns closely with real-world discourse analysis needs.Hopefully,this proposed framework can provide a robust foundation for future advancements in discourse relation models and can be potentially applied in areas that require nuanced language comprehension.

关 键 词：隐式篇章关系识别 序列到序列模型 语句角色标注增强 连接词生成 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]