融合标签特征和胶囊注意力的口语理解方法

Spoken language understanding method fusing label features and capsule attention

针对目前意图检测和槽位填充联合学习中未充分考虑交互前标签特征信息的有效提取和融合,缺乏对交互后标签特征的提炼问题,提出一种融合标签特征和胶囊注意力的口语理解方法。主要由意图与槽位标签特征融合交互(label feature fusion interactive, LFFI)和多头胶囊注意力机制(multi-head capsule attention, MHCA)两大关键模组组成。LFFI-MHCA通过LFFI提取序列中有效的意图和槽位标签信息,对两者进行融合和交互;利用MHCA对交互过程中产生的不同子空间信息进行提炼,获得更为精确的意图和槽位标签特征。该模型在ATIS和SNIPS数据集上进行实验,句子准确率分别为88.1%和89.0%,验证了该模型的有效性。

In view of the current joint learning of intent detection and slot filling,the effective extraction and fusion of label feature information before interaction is not fully considered,and the extraction of label features after interaction is lacking,a spoken language understanding was proposed by fusing label features and capsule attention.The method mainly consisted of two key modules,intent and slot label feature fusion interactive(LFFI)and multi-head capsule attention(MHCA).LFFI-MHCA was used to extract the effective intent and slot label information in the sequence through LFFI and further the extraction was fused and interacted.To better obtain more accurate intent and slot label features,MHCA was applied to refine the different subspace information generated during the interaction process.The model was tested on the ATIS and SNIPS datasets,and the sentence accuracy rates are 88.1%and 89.0%,respectively,verifying the effectiveness of the model.