基于多视角注意力的答案选择算法研究
首发时间:2020-03-18
摘要:答案选择是自然语言处理领域中的一个重要子任务,同时也是自动问答系统的一个极其重要的支撑技术。由于答案选择任务主要解决问题和答案之间的相关性匹配,而注意力机制可以提供灵活并有效的信息交互与利用的方式,继而成为问答系统中不可或缺的一个关键技术模块。本文提出一种基于多视角注意力机制的答案选择算法,通过多种注意力类型(协同注意力、自注意力)和多种注意力变体(最大池化、平均池化、软对齐)的调用来建模多角度的语义视图,从而提高语义编码的完整性和准确性。同时为了消除同时执行多种注意力机制所需架构工程的昂贵需求,提升算法的计算效率,本文提出将注意力作为一种特征增强方式使用,实现多种注意力机制的可扩展调用。通过压缩函数返回标量特征,并将特征重新附加到原始的单词表示上,为后续编码层提供包含句子内部的知识和句子之间的知识的特征,改进表示学习过程。模型在事实型问答数据集(TrecQA)、开放域数据集(WikiQA)和社区问答数据集(SemEval-2016 CQA和YahooCQA)上进行实验,均实现了目前最好的性能。通过消融研究,也证明了多视角注意力机制的有效性。
For information in English, please click here
A Multi-view Attention Network for Answer Selection Algorithm
Abstract:Answer selection is an important sub-task in the field of natural language processing, and it is also an extremely important supporting technology for automatic question answering systems. Since answer selection task mainly solves the semantic matching between question and answer, the attention mechanism could provide an effective way of information interaction, thus becoming an indispensable key technical module in the question answering system. This paper proposes a multi-view attention network which uses multiple attention types(co-attention and self-attention) and multiple attention variants(max pooling, average pooling, soft alignment) to model multi-perspective semantic views, thus improving the completeness and accuracy of 40 semantic encoding. At the same time, in order to eliminate the expensive requirements of architectural engineering and improve the computational efficiency of the algorithm, this paper proposes to re-imagine attention as a form of feature argumentation method, achieving multiple attention casts. The model returns scalar feature using compressed function after soft attention operations, and re-attach it to the original word representation, providing hints with global knowledge and cross-sentence 45 knowledge for subsequent encoding layers, which could improve representation learning. Experiments on the factual based question answering dataset (TrecQA), open-domain dataset (WikiQA), and community question answering dataset (SemEval-2016 CQA and YahooCQA) outperform existing state-of-the-art models and ablation studies prove the effectiveness of the multi-view attention mechanism.
Keywords: artificial intelligence question answering system answer selection attention mechanism
基金:
引用
No.****
同行评议
勘误表
基于多视角注意力的答案选择算法研究
评论
全部评论0/1000