Domain-specific Answer Sentence Selection with Terminology Augmentation and Cascade Attention

Xiaotong Lyu, Denghao Ma, Yueguo Chen

Abstract: The online consulting service has become a popular and convenient channel for people to seek professional replies. Since the replies are often lengthy, the answer highlighting is critical for users to identify the core answers. So we study the task of \(\overline{A}\)nswer \(\overline{S}\)entence \(\overline{S}\)election in specific \(\overline{d}\)omains (ASSD), which is to select core answer sentences in replies for highlighting. Even pre-trained language models (PLMs) have made great progress in ASSD, there is still a significant untapped potential in deep understanding domain-specific texts which are replete with specialized terminologies. So we propose a novel Terminology-augmented Cascade Attention(TACA) framework by incorporating domain-specific knowledge into PLMs, to achieve better text understanding and then accomplish the ASSD task more effectively. In the framework, we first design the terminology-augmented multi-channel semantic model to deeply mine the semantics of both questions and answers. Second, the cascade attention mechanism is proposed to incorporate multi-channel semantics and achieve fine-grained semantic matching between questions and answers. Extensive experiments on two datasets show that TACA significantly improves the accuracy in the ASSD task.