Enhanced LSTM for Natural Language Inference

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1657–1668 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1152 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1657–1668 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1152 Enhanced LSTM for Natural Language Inference Qian Chen University of Science and Technology of China cq1231@mail.ustc.edu.cn Xiaodan Zhu National Research Council Canada xiaodan.zhu@nrc-cnrc.gc.ca Zhenhua Ling University of Science and Technology of China zhling@ustc.edu.cn Si Wei iFLYTEK Research siwei@iflytek.com Hui Jiang York University hj@cse.yorku.ca Diana Inkpen University of Ottawa diana@site.uottawa.ca Abstract Reasoning and inference are central to hu- man and artificial intelligence. Modeling inference in human language is very chal- lenging. With the availability of large an- notated data (Bowman et al., 2015), it has recently become feasible to train neural network based inference models, which have shown to be very effective. In this paper, we present a new state-of-the-art re- sult, achieving the accuracy of 88.6% on the Stanford Natural Language Inference Dataset. Unlike the previous top models that use very complicated network architec- tures, we first demonstrate that carefully de- signing sequential inference models based on chain LSTMs can outperform all previ- ous models. Based on this, we further show that by explicitly considering recursive ar- chitectures in both local inference model- ing and inference composition, we achieve additional improvement. Particularly, in- corporating syntactic parsing information contributes to our best result—it further im- proves the performance even when added to the already very strong model. 1 Introduction Reasoning and inference are central to both human and artificial intelligence. Modeling inference in human language is notoriously challenging but is a basic problem towards true natural language un- derstanding, as pointed out by MacCartney and Manning (2008), “a necessary (if not sufficient) condition for true natural language understanding is a mastery of open-domain natural language in- ference.” The previous work has included extensive research on recognizing textual entailment. Specifically, natural language inference (NLI) is concerned with determining whether a natural- language hypothesis h can be inferred from a premise p, as depicted in the following example from MacCartney (2009), where the hypothesis is regarded to be entailed from the premise. p: Several airlines polled saw costs grow more than expected, even after adjusting for inflation. h: Some of the companies in the poll reported cost increases. The most recent years have seen advances in modeling natural language inference. An impor- tant contribution is the creation of a much larger annotated dataset, the Stanford Natural Language Inference (SNLI) dataset (Bowman et al., 2015). The corpus has 570,000 human-written English sentence pairs manually labeled by multiple human subjects. This makes it feasible to train more com- plex inference models. Neural network models, which often need relatively large annotated data to estimate their parameters, have shown to achieve the state of the art on SNLI (Bowman et al., 2015, 2016; Munkhdalai and Yu, 2016b; Parikh et al., 2016; …

Content truncated to 500 words