基于双图注意力的多领域口语语言理解联合模型  

DualGAT Based Joint Model for Multi-Domain Spoken Language Understanding

作　　者：贾旭 彭敏[1] JIA Xu;PENG Min(School of Computer Science,Wuhan University,Wuhan,Hubei 430072,China)

机构地区：[1]武汉大学计算机学院,湖北武汉430072

出　　处：《中文信息学报》2023年第10期76-85,共10页Journal of Chinese Information Processing

基　　金：科技创新2030-“新一代人工智能”重大项目(2021ZD0113304);国家自然科学基金(62072346);湖北省重点研发计划项目(2021BBA099、2021BBA029)。

摘　　要：多领域口语语言理解包括多意图识别和槽填充两个子任务,现有研究通过构建语句中的意图和槽之间的关联提升模型的表现。然而现有研究将多领域场景下的意图和槽看作相互独立的标签,忽视了标签之间领域内和领域间的结构关联。该文提出细粒度标签图和领域相关图的双图注意力联合模型。具体来说,细粒度标签图将意图和槽标签分成细粒度分片,建模分片之间的结构性关联和上下文表示的语义特征。领域相关图通过标签间的领域信息,建模预测意图和对应领域内槽的关联,减少图中的冗余关联。实验结果表明,在两个公开的数据集上,该文提出的模型均优于基准模型。Multi-domain spoken language understanding consists of two subtasks:multi-intent detection and slot filling.Previous studies have achieved notable performance by establishing correlations between intents and slots,though failing to capture the inherent intra-domain and inter-domain structural correlations between intent and slot labels.This paper introduces a novel joint model based on DualGAT,incorporating the fine-grained label graph and the domain-related graph.The fine-grained label graphs split the intent and slot labels into fine-grained pieces,capturing the structural correlations between the pieces and the semantic features of the contextual representation.The domain-related graph leverages domain information to model correlations between predicted intents and their corresponding slots,thereby reducing redundant correlations in the graph.Experimental results show that our model outperforms the baselines on two publicly available datasets.

关 键 词：多领域口语语言理解 多意图识别 细粒度标签图 领域相关图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]