A Noise-Removal of Knowledge Graph Framework for Profile-Based Spoken Language Understanding

Leyi Lao, Peijie Huang, Zhanbiao Zhu, Hanlin Liu, Peiyi Lian, Yuhong Xu

Abstract: To alleviate the ambiguity of user utterances, a profile-based spoken language understanding (PROSLU) task has been proposed recently, which uses supporting profile information as supplementary knowledge. The knowledge graph, as one of the supporting profile information, contains many entities with rich attribute information which do help the performance of SLU. However, existing models treat all entities as useful, resulting in a lot of noise being fed. Moreover, not all the attribute information of the entity is necessary, and there is a lot of noise and redundancy. In this paper, we propose a Noise-Removal of Knowledge Graph framework for PROSLU (NRKG-PROSLU), with two different kinds of denoising. One is a hard denoising by entity selection, where we introduce a small clean dataset and propose an auxiliary model, BERT-based entity selection, to filter out entities irrelevant to user utterances, namely noisy entities. The other is a soft denoising by entity attribute information selection, where we introduce keywords-based local semantic selection, which uses keywords to give more weight to relevant local semantics, so as to capture task-related information in the selected entities for reducing noise and redundancy. Then, the denoised knowledge graph is used to assist the SLU model, which achieves better performance than competitive models on the public PROSLU dataset.