基于多层次动态门控推理网络的文本蕴含识别

Multilevel dynamic gated inference network for recognizing textual entailment

摘要:

现有的文本蕴含模型通常计算一次词级别注意力得到两段文本在不同层面的交互特征，但对于文本不同层面的理解，不同重要词的注意力应该是不同的，并且一次词级注意力推理仅能捕捉到文本对局部特征.针对这个问题，提出一种多层次动态门控推理网络(MDGIN)，该网络结合了词级别信息的细粒度推理和句子级别门控机制来动态捕捉文本对的语义信息，并采用不同注意力计算方式提取文本对不同层面的语义特征，共同推理文本对的蕴含关系.本文在两个文本蕴含数据集上均做了实验，相较于基准模型和现有主流模型，准确率提升了0.4%-1.7%，通过消融分析，进一步验证了本文模型各部分结构的有效性.

Abstract:

Most existing models of recognizing textual entailment (RTE) get the interaction features between a premise and a hypothesis by an attention matrix at word level. However, the attention of important words should be different with the degree of understanding from diverse aspects, and only the local features are captured. To solve the problem above, the model with Multilevel Dynamic Gated Inference Network (MDGIN) is proposed, which combines the finegrained wordlevel information and sentencelevel gating mechanism to dynamically capture the relationships of text pairs. Moreover, the model extracts the different semantic features by diverse attention ways. In this paper, experiments are carried out on two textual datasets. Compared with the benchmark models and the existing mainstream models, the accuracy is improved by 0.4%~1.7%. The effectiveness of each part of the model is further verified by ablation analysis.