DR-BiLSTM: Dependent Reading Bidirectional LSTM for Natural Language Inference

Proceedings of NAACL-HLT 2018, pages 1460–1469 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics DR-BiLSTM: Dependent Reading Bidirectional LSTM for Natural Language Inference∗ Reza Ghaeini1†, Sadid A. Hasan2, Vivek Datla2, Joey Liu2, Kathy Lee2, Ashequl Qadir2, Yuan Ling2, Aaditya Prakash2, Xiaoli Z. Fern1 and Oladimeji Farri2 1Oregon State University, Corvallis, OR, USA 2Artificial Intelligence Laboratory, Philips Research North America, Cambridge, MA, USA {ghaeinim,xfern}@eecs.oregonstate.edu {sadid.hasan,vivek.datla,joey.liu,kathy.lee 1,ashequl.qadir}@philips.com {yuan.ling,aaditya.prakash,dimeji.farri}@philips.com Abstract We present a novel deep learning archi- tecture to address the natural language in- ference (NLI) task. Existing approaches mostly rely on simple reading mecha- nisms for independent encoding of the premise and hypothesis. Instead, we pro- pose a novel dependent reading bidirec- tional LSTM network (DR-BiLSTM) to efficiently model the relationship between a premise and a hypothesis during encod- ing and inference. We also introduce a sophisticated ensemble strategy to com- bine our proposed models, which notice- ably improves final predictions. Finally, we demonstrate how the results can be improved further with an additional pre- processing step. Our evaluation shows that DR-BiLSTM obtains the best single model and ensemble model results achiev- ing the new state-of-the-art scores on the Stanford NLI dataset. 1 Introduction Natural Language Inference (NLI; a.k.a. Recog- nizing Textual Entailment, or RTE) is an important and challenging task for natural language under- standing (MacCartney and Manning, 2008). The goal of NLI is to identify the logical relationship (entailment, neutral, or contradiction) between a premise and a corresponding hypothesis. Table 1 shows few example relationships from the Stan- ford Natural Language Inference (SNLI) dataset (Bowman et al., 2015). Recently, NLI has received a lot of attention from the researchers, especially due to the avail- ∗ ArXiv version of this work can be found here (arxiv.org/pdf/1802.05577.pdf). † This work was conducted as part of an internship pro- gram at Philips Research. Pa A senior is waiting at the Relationship window of a restaurant that serves sandwiches. Hb A person waits to be Entailment served his food. A man is looking to order Neutral a grilled cheese sandwich. A man is waiting in line Contradiction for the bus. aP, Premise. bH, Hypothesis. Table 1: Examples from the SNLI dataset. ability of large annotated datasets like SNLI (Bow- man et al., 2015). Various deep learning models have been proposed that achieve successful results for this task (Gong et al., 2017; Wang et al., 2017; Chen et al., 2017; Yu and Munkhdalai, 2017a; Parikh et al., 2016; Zhao et al., 2016; Sha et al., 2016). Most of these existing NLI models use at- tention mechanism to jointly interpret and align the premise and hypothesis. Such models use sim- ple reading mechanisms to encode the premise and hypothesis independently. However, such a com- plex task require explicit modeling of dependency relationships between the premise and the hypoth- esis during the encoding and inference processes to prevent the network from the loss of relevant, contextual information. In this paper, we refer to such strategies as dependent reading. There are some alternative …

Content truncated to 500 words