TDGCN:触发器增强的两阶段动态图卷积网络的对话关系抽取研究
TDGCN:Research on Conversation Relationship Extraction of Two Stage Dynamic Graph Convolutional Networks Enhanced by Triggers
自彦丞 1李卫疆2
作者信息
- 1. 昆明理工大学信息工程与自动化学院,昆明 650500
- 2. 昆明理工大学云南省人工智能重点实验室,昆明 650500
- 折叠
摘要
随着互联网中对话数据的不断增加,从中提取关系三元组对于自然语言处理的各个下游任务至关重要.为了改进对话关系抽取的性能,D.Yu等人在数据集中引入了"触发器"的概念,该概念为关系抽取提供了重要的线索.然而,目前对于触发器的应用还仅仅限于将其作为一个模型训练的附加任务,并未在关系三元组推理中充分利用.本文提出了一个两阶段的动态图模型,通过引入动态机制,有效地改进了现有静态构造的图注意力模型在处理关系重叠时的歧义问题.并且在动态图模型中引入了触发器节点,以便更充分地利用触发器来进行关系推理.整个模型在DialogRE数据集上进行了实验,相对于基线模型,该模型在验证集上的F1值提升了 2.2％,在测试集上提升了 2％.并且本文对所提出的机制进行了进一步分析,通过实验验证了其有效性.
Abstract
With the continuous increase of conversation data in the Internet,extracting relational triples from it is crucial for various downstream tasks of natural language processing.In order to improve the performance of dialogue relationship extraction,D.Yu et al.introduced the concept of"triggers"in the dataset,which provides important clues for relationship extraction.However,the current ap-plication of triggers is only limited to using them as an additional task for model training,and has not been fully utilized in relational triplet reasoning.This article proposes a two-stage dynamic graph model,which effectively improves the ambiguity problem of existing statically constructed graph attention models when dealing with overlapping relationships by introducing dynamic mechanisms.And trigger nodes are introduced in the dynamic graph model to more fully utilize triggers for relational reasoning.The entire model was tested on the DialogRE dataset,and compared to the baseline model,the F1 value of the model increased by 2.2％on the validation set and 2％on the test set.And this article further analyzes the proposed mechanism and verifies its effectiveness through experiments.
关键词
动态图注意力网络/对话关系抽取/触发器Key words
dynamic graph attention network/dialogue relationship extraction/trigger引用本文复制引用
出版年
2025
NETL
NSTL
万方数据