Using Shallow Semantic Parsing and Relation Extraction for Finding Contradiction in Text

International Joint Conference on Natural Language Processing, pages 1017–1021, Nagoya, Japan, 14-18 October 2013. Using Shallow Semantic Parsing and Relation Extraction for Finding Contradiction in Text Minh Quang Nhat Pham, Minh Le Nguyen and Akira Shimazu Japan Advanced Institute of Science and Technology 1-1 Asahidai, Nomi, Ishikawa, 923-1292, JAPAN {minhpqn,nguyenml,shimazu}@jaist.ac.jp Abstract Finding contradiction text is a fundamen- tal problem in natural language under- standing. Previous work on finding contra- diction in text incorporate information de- rived from predicate-argument structures as features in supervised machine learning frameworks. In contrast to previous work, we combine shallow semantic representa- tions derived from semantic role labeling with binary relations extracted from sen- tences in a rule-based framework. Eval- uation experiments conducted on stan- dard data sets indicated that our system achieves better recall and F1 score for con- tradiction detection than most of baseline methods, and the same recall as a state of the art supervised method for the task. 1 Introduction Contradiction detection (CD) in text is a funda- mental task in natural language understanding, and necessary for many applications (De Marn- effe et al., 2008; Voorhees, 2008). For instance, contradictions need to be recognized by question answering systems or multi-document summariza- tion systems (Harabagiu et al., 2006). The task is to detect whether the contradiction relationship exists in a pair of a text T and a hypothesis H. There are several approaches to the CD task. Contradiction detection can be formalized as a binary classification problem (Harabagiu et al., 2006; De Marneffe et al., 2008). The main effort of work which adopt this approach is to find out ef- fective features for recognizing contradiction. The other approach is using functional relations indi- cated by verb or noun phrases for detecting con- tradiction (Ritter et al., 2008). Beyond string-based matching approaches, one can approach to the CD task by applying logical inference techniques. Although the logical infer- ence approach may obtain good precision, it is not widely used for the task due to the fact that full predicate-logic analysis is currently not prac- tical for wide-coverage semantic processing (Bur- chardt et al., 2009). Given that fact, (Burchardt et al., 2009) pointed out that using shallow seman- tic representations based on predicate-argument structures and frame knowledge is an intuitive and straightforward approach to textual inference tasks. In contrast to previous work which integrate predicate-argument structures as features in ma- chine learning-based systems (Harabagiu et al., 2006; De Marneffe et al., 2008), this paper com- bines shallow semantic representations derived from semantic role labeling with binary relations extracted from sentences for the CD task. The proposed system consists of two modules. The first module relies on the alignment of semantic role (SRL) frames extracted from the text and the hypothesis in each pair while the second one per- forms contradiction detection over binary relations extracted from the pair. If the SRL-based mod- ule fails to identify the contradiction relationship in the pair, the second module will be applied. We expect that the second module will …

Content truncated to 500 words