Embedding WordNet Knowledge for Textual Entailment

Proceedings of the 27th International Conference on Computational Linguistics, pages 270–281 Santa Fe, New Mexico, USA, August 20-26, 2018. 270 Embedding WordNet Knowledge for Textual Entailment Yunshi Lan School of Information Systems Singapore Management University yslan.2015@phdis.smu.edu.sg Jing Jiang School of Information Systems Singapore Management University jingjiang@smu.edu.sg Abstract In this paper, we study how we can improve a deep learning approach to textual entailment by incorporating lexical entailment relations from WordNet. Our idea is to embed the lexical entail- ment knowledge contained in WordNet in specially-learned word vectors, which we call “entail- ment vectors.” We present a standard neural network model and a novel set-theoretic model to learn these entailment vectors from word pairs with known lexical entailment relations derived from WordNet. We further incorporate these entailment vectors into a decomposable attention model for textual entailment and evaluate the model on the SICK and the SNLI dataset. We find that using these special entailment word vectors, we can significantly improve the performance of textual entailment compared with a baseline that uses only standard word2vec vectors. The final performance of our model is close to or above the state of the art, but our method does not rely on any manually-crafted rules or extensive syntactic features. 1 Introduction Recognizing textual entailment (RTE) is the task of determining whether a hypothesis sentence can be inferred from a given premise sentence. The task has been well studied since it was first introduced by Dagan et al. (2006). Recently, there has been much interest in applying deep learning models to RTE (Bowman et al., 2015; Rocktaschel et al., 2016; Wang and Jiang, 2016; Parikh et al., 2016; Sha et al., 2016). These models usually do not perform any linguistic analysis or require any feature engineering but have been shown to perform very well on this task. Intuitively, lexical-level entailment relations should help sentence-level RTE. For example, Table 1 shows a premise and a hypothesis taken from the test data of the SICK dataset (Marelli et al., 2014). We can see that in this example the premise entails the hypothesis, and in order to correctly identify this relation, one has to know that the word kettle entails the word pot. However, if we train a neural network model on a set of labeled sentence pairs, and if the training dataset does not contain the word pair kettle and pot anywhere, it would be hard for the learned model to know that kettle entails pot and subsequently predict the relation between the premise and the hypothesis to be entailment. On the other hand, from WordNet we can easily find out that pot is a direct hypernym of kettle and therefore kettle should entail pot. If this kind of prior knowledge can be injected into a trained RTE model, then for sentence pairs with words that do not occur in the training data but can be found in WordNet, the RTE predictions could potentially be made easier. Premise: Someone is stirring chili in a kettle. Hypothesis: Someone is …

Content truncated to 500 words