Asynchronous Deep Interaction Network for Natural Language Inference

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2692–2700, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2692 Asynchronous Deep Interaction Network for Natural Language Inference Di Liang∗, Fubao Zhang∗, Qi Zhang†, and Xuanjing Huang School of Computer Science, Shanghai Key Laboratory of Intelligent Information Processing Fudan University, Shanghai, P.R.China 201203 {liangd17, fbzhang17, qz, xjhuang}@fudan.edu.cn Abstract Natural language inference aims to predict whether a premise sentence can infer another hypothesis sentence. Existing methods typi- cally have framed the reasoning problem as a semantic matching task. The both sentences are encoded and interacted symmetrically and in parallel. However, in the process of reason- ing, the role of the two sentences is obviously different, and the sentence pairs for NLI are asymmetrical corpora. In this paper, we pro- pose an asynchronous deep interaction net- work (ADIN) to complete the task. ADIN is a neural network structure stacked with mul- tiple inference sub-layers, and each sub-layer consists of two local inference modules in an asymmetrical manner. Different from previous methods, this model deconstructs the reason- ing process and implements the asynchronous and multi-step reasoning. Experiment results show that ADIN achieves competitive perfor- mance and outperforms strong baselines on three popular benchmarks: SNLI, MultiNLI, and SciTail. 1 Introduction Natural language inference (NLI) is a pivotal and fundamental task in natural language understand- ing and artificial intelligence. The goal of NLI is to predict whether a premise sentence can in- fer another hypothesis sentence. As illustrated in Table 1, logical relationships between the two sen- tences include entailment (if the premise is true, then the hypothesis must be true), contradiction (if the premise is true, then the hypothesis must be false), and neutral (neither entailment nor contra- diction). As a core task, conventional approaches have studied various aspects of the inference prob- ∗Equal contribution. Alphabetical order of the last name. †Corresponding author. Premise: A dog is jumping for a Frisbee in the snow. Hypothesis: An animal is playing with a plastic toy. Label: Entailment Premise: He was crying like his mother had just walloped him. Hypothesis: He was crying like his mother hit him with a spoon. Label: Neutral Premise: Several men in front of a white building. Hypothesis:Several people in front of a gray building. Label: contradiction Table 1: Examples of natural language inference. lem (MacCartney and Manning, 2008; Heilman and Smith, 2010). Thanks to the release of the largest publicly available corpus - the Stanford Natural Language Inference (SNLI) corpus (Bow- man et al., 2015), neural network-based models have also been successfully used for this task (Parikh et al., 2016; Chen et al., 2016; Tay et al., 2018; Duan et al., 2018). These methods typically treat the premise sentence and the hypothesis sen- tence equally, learn an alignment of sub-phrases in both sentences symmetrically and in parallel, and fuse local information for making a global deci- sion at the sentence level. They all frame the …

Content truncated to 500 words