一种端到端的生成式问答模型
首发时间:2019-01-28
摘要:问答系统是自然语言处理领域重要的研究方向之一,近年来受到人们越来越多的关注。本文基于双向注意力机制和"拷贝-生成"机制,提出了一种新颖的端到端生成式问答模型,融合答案抽取和生成两个过程,并引入Coverage机制缓解生成重复的问题。相比于经典的问答模型,端到端融合模型一方面可以抽取段落实体、未登录词等关键信息,另一方面又能够生成自然语言形式的答案。在MSMARCO v2.1数据集上进行了实验,端到端融合模型在Rouge-L和BLEU-n等指标上均优于纯生成模型和纯抽取模型。
For information in English, please click here
An End-to-End Generated Question and Answer Model
Abstract:The question and answer system is one of the important research directions in the field of natural language processing, and has received more and more attention in recent years. Based on the two-way attention mechanism and the "copy-generation" mechanism, this paper proposes a novel end-to-end generated question-and-answer model, which integrates the answer extraction and generation processes, and introduces the Coverage mechanism to alleviate the problem of generating duplicates. Compared with the classic question-and-answer model, the end-to-end fusion model can extract key information such as paragraph entities and unregistered words on the one hand, and generate answers in natural language form on the other hand. Experiments were carried out on the MS MARCO v2.1 dataset. The end-to-end fusion model is superior to the purely generated model and the purely extracted model in terms of Rouge-L and BLEU-n.
Keywords: Intelligent Science and Technology Natural language generation Machine reading comprehension
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
一种端到端的生成式问答模型
评论
全部评论