Abstract: Knowledge Graph（KG） is one of the key technologies for implementing an intelligent Question Answering system（QA）. It can reduce customer service costs and enhance their self-service capabilities. It has a lot of commercial values and research meanings. To reduce the fuzzy of Chinese language questions and the high cost of online service operation and maintenance in a KG based QA, a smart grid customer service Question Answering system（HDKG-QA） is proposed. It is based on the hybrid domain feature of KG. It first identifies the entity based on a LSTM model. Then it proposes a semantic enhancement method based on the topic comparison to accurately find the external knowledge. It uses heuristic rules to get the optimal answer. Periodically, it updates the global KG according to the result of ILP solver. HDKG-QA can achieve high entity/predication recognition accuracy. It automatically maps domain knowledge to the local KG, and updates online KG. It can achieve high quality response with low response delay. The system is verified by the actual Q&A dataset in the State Grid Chongqing Electric Power Corporation, the experimental results show that the accuracy of the QA is improved by 17%. Through introducing LSTM and semantic enhancement methods, the heuristic rules and the sorting strategy can increase the accuracy by 8%. The Q&A response latency can be reduced by 9% with the same accuracy by introducing the ILP solver.

Key words: knowledge graph, question answering system, mapping model of domain features, knowledge quality awareness

摘要： 知识图谱（KG）是实现领域问答系统的关键技术之一，能够降低客服成本，推进客户自助服务的智能化，具有较大的商用价值和研究意义。针对基于KG问答系统中存在的中文问题表达模糊、线上服务运维成本高的问题，融合领域特征知识图谱的电网客服问答系统（HDKG-QA），其能基于LSTM模型识别实体/断言，基于主题比较的语义增强方法准确寻找外部知识，使用启发式规则优化答案候选集，并定期根据ILP求解器设置全局KG的更新策略。HDKG-QA能够达到较高的实体/断言识别准确率，自动将领域知识映射为本地KG，快速实现服务知识库的在线更新，达到以较低的响应延迟实现高准确率的回答。根据国网重庆市电力公司信息通信分公司的实际客服问答数据集对本系统进行验证，实验结果表明通过引入LSTM和语义增强方法，问答系统的准确率提高了17%；基于启发式规则的优化答案排序策略将准确率提高了8%；通过引入ILP求解器，在保障同样准确率的情况下，问答响应延迟降低了9%。

关键词: 知识图谱, 问答系统, 领域知识映射, 知识质量感知