基于全局知识和前后向序列学习的自然答案生成模型

Natural answer generation with global knowledge and backward & forward sequence-to-sequence learning

传统问答系统只能返回孤零零的实体或短语作为答案,这种答案不是一个自然友好的回复形式,事实上,大多数用户希望问答系统的答案不仅需要内容正确,也需要形式自然。利用联合消岐模型把从社区网站获取的自然问题—答案对与知识库进行对齐,研究问答系统中事实类问题的自然答案生成模型。基于知识库的全局表示学习,提出了一个两阶段的自然答案生成模型,第一阶段使用了各种神经网络模型匹配问题和事实,第二阶段利用前后向序列学习模型生成自然语言回复形式的答案。在开放数据集上的实验结果表明,相较于基准系统,该模型不仅能够提升答案准确率14. 3%,也能够生成更加自然的答案回复。

The traditional question answering systems merely output a solitary entity/phrase for each question,rather than a natural and friendly reply.In fact,most existing QA related products need to generate not only correct but also natural answering responses.In this paper,we devote to generating natural answers for factual questions.We first obtain the training question-answer pairs from community website and grounded them with a knowledge base(KB)using a joint disambiguation model.Then we propose a two-stage natural answer generation model with encoded global knowledge.A variety of neural models are used to match the question with facts of KB in the first stage.Lastly,we utilize a backward and forward sequence-to-sequence model to generate natural responses based on input questions and matched facts.Our empirical study on a public dataset demonstrates that the proposed model yields 14.3%accuracy gain over the state-of-the-art baseline,and it also enables to generate natural responses based on a manual evaluation.