A Hybrid Approach of Deep Semantic Matching and Deep Rank for Context Aware Question Answer System

All authors contributed equally to this manuscript. A Hybrid Approach of Deep Semantic Matching and Deep Rank for Context Aware Question Answer System Shu-Yi Xie, Chia-Hao Chang†, Zhi Zhang, Yang Mo, Lian-Xin Jiang, Yu-Sheng Huang, Jian-Ping Shen AI Department Ping An Life Insurance of China, Ltd. {xieshuyi542,zhangjiahao206†,zhangzhi600,moyang853,jianglianxin769, huangyusheng112,shenjianping324}@pingan.com.cn Correspondence†:strategist922@gmail.com Abstract Most of the existing Question Answer Systems focused on searching answers from the Knowledge-Base (KB), and ignore context aware information. Many Question Answer models perform well on public data-sets, but too complicated to be efficient in real world cases. Effectiveness, concurrency and system availability are equally important in industry which have large data and requests, we propose a Context Aware Question Answer System based on the Information Retrieval with Deep Semantic Matching and Deep Rank. It has been applied to the online question answer system for insurance Question Answer. By these means, we achieve both high QPS (Query Per Second) and effectiveness. Our approach improves the system’s ability to understand the question with context aware coreference resolution, subject completion, and the long sentence compression. After the matching questions are recalled from the ElasticSearch, Siamese CBOW (Continues Bag-Of-Words Model) and KBQA filter some unreasonable ones by entity alignment. After the result is sorted by the deep rank model with co-occurrence words and semantic features, our system does clarification or answer output. Finally, for those questions that we are unable to provide answers, a dialogue mining module as part of our Smart Knowledge-Base Platform is developed. This results in more than 10 times improvement in terms of efficiency for manpower involved in data labeling process. KEYWORDS Question Answering, Coreference Resolution, Error Correction, Sentence Compression, Deep Semantic Match, Deep Rank, Knowledge-Base Management, Insurance Domain The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 72-81 ©The Association for Computational Linguistics and Chinese Language Processing 72 . 1 INTRODUCTION The question answer system has been widely used in intelligent customer service, personal assistants, and dialogue robots. In 2018, the pretrain techniques based on a massive corpus pre- training model have made breakthroughs in multiple NLP tasks including Semantic Match. Representative models are Elmo[9], GPT[10], BERT[8]. Higher accuracy, compared with the Siamese CBOW, can be achieved by fine-tuning BERT on downstream tasks, but the model makes inference time much longer, the running efficiency does not meet the requirements of our online products. We propose a high-efficiency contextual referential solution based on syntax analysis to solve the problems of subject missing and pronoun resolution in the question- and-answer scenario in insurance industry that achieved good results. The voice input brings convenience to users but at the same time introduces typos in the results after the text processing. We use the insurance specific noun dictionary with the error correction model of Transformer[7] to improve the input from ASR. For the purpose of increasing the accuracy of matching sentences of the user’s input with terms from Knowledge-Base, we use an efficient sentence compression algorithm, which can filter some insignificant content and retain some core content of …

Content truncated to 500 words