Spoken language understanding (SLU) is an essential component inconversational systems. Most SLU component treats each utterance independently,and then the following components aggregate the multi-turn information in theseparate phases. In order to avoid error propagation and effectively utilizecontexts, prior work leveraged history for contextual SLU. However, theprevious model only paid attention to the content in history utterances withoutconsidering their temporal information and speaker roles. In the dialogues, themost recent utterances should be more important than the least recent ones.Furthermore, users usually pay attention to 1) self history for reasoning and2) others' utterances for listening, the speaker of the utterances may providesinformative cues to help understanding. Therefore, this paper proposes anattention-based network that additionally leverages temporal information andspeaker role for better SLU, where the attention to contexts and speaker rolescan be automatically learned in an end-to-end manner. The experiments on thebenchmark Dialogue State Tracking Challenge 4 (DSTC4) dataset show that thetime-aware dynamic role attention networks significantly improve theunderstanding performance.
展开▼
