Enabling Transitivity for Lexical Inference on Chinese Verbs Using Probabilistic Soft Logic

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 110–119, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Enabling Transitivity for Lexical Inference on Chinese Verbs Using Probabilistic Soft Logic Wei-Chung Wang, Lun-Wei Ku Academia Sinica 128 Academia Road, Section2 Nankang, Taipei 11529, Taiwan {anthonywang, lwku}@iis.sinica.edu.tw Abstract To learn more knowledge, enabling tran- sitivity is a vital step for lexical infer- ence. However, most of the lexical infer- ence models with good performance are for nouns or noun phrases, which can- not be directly applied to the inference on events or states. In this paper, we construct the largest Chinese verb lexical inference dataset containing 18,029 verb pairs, where for each pair one of four in- ference relations are annotated. We fur- ther build a probabilistic soft logic (PSL) model to infer verb lexicons using the logic language. With PSL, we easily en- able transitivity in two layers, the observed layer and the feature layer, which are in- cluded in the knowledge base. We further discuss the effect of transitives within and between these layers. Results show the performance of the proposed PSL model can be improved at least 3.5% (relative) when the transitivity is enabled. Fur- thermore, experiments show that enabling transitivity in the observed layer benefits the most. 1 Introduction Lexical inference is an important component of natural language understanding for NLP tasks such as textual entailment (Garrette et al., 2011), metaphor detection (Mohler et al., 2013), and text generation (Biran and McKeown, 2013) to ac- quire implications not explicitly written in context. Given two words, the goal of lexical inferences is to detect whether there is an inference relation between the lexicon pair. For example, the word ‘buy’ entails the word ‘have’. With the help of lex- ical inference system, we can know “Mom has ap- ples” from the ground truth “Mom buys apples”to answer the question “Who has apples?” without explicitly mentioning it. An intuitive solution to this problem is to first represent the sense of words in the lexicon to cal- culate the confidence of inferences from one sense to another, or to build a classifier to distinguish inference relations from other relations. Most re- lated research is of one of these two types (Szpek- tor and Dagan, 2008a; Kiela et al., 2015). How- ever, for this problem it is difficult for these mod- els to take into account transitivity. In the frame- work of a lexical inference system, transitivity can be included in three layers: the observed layer, the feature layer, and the prediction layer. Figure 1 illustrates these layers and the corresponding tran- sitives. The observed layer includes inference re- lations we already know, e.g., true inferences from the gold labels or ontologies; the feature layer in- cludes the observed features for all lexicon pairs to be predicted,i.e.,features for the testing data, and the predicted layer saves the predicted inference pairs, i.e., the relations of pairs in the testing data, predicted by the model. As inference usually …

Content truncated to 500 words