Unsupervised Verb Inference from Nouns Crossing Root Boundary

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1248–1259, Dublin, Ireland, August 23-29 2014. Unsupervised Verb Inference from Nouns Crossing Root Boundary Soon Gill Hong Department of KSE KAIST Daejeon, Republic of Korea hsoongil@gmail.com Sin-Hee Cho Department of KSE KAIST Daejeon, Republic of Korea chosinhee@kaist.ac.kr Mun Yong Yi Department of KSE KAIST Daejeon, Republic of Korea munyi@kaist.ac.kr Abstract Inference about whether a word in one text has similar meaning to another word in the other text is an essential task in order to understand whether two texts have similar meaning. However, this inference becomes difficult especially when two words do not share a lexical root, do not have the same argument structure, or do not have the same part-of-speech. This paper presents an unsupervised approach for inferring verbs from nouns along with a new online resource PreDic (PREdicate DICtionary) that contains verbs inferred from nouns sharing similar concepts but not the root. The verbs in PreDic are categorized into three groups, enabling applications to target precision-oriented, recall-oriented, or harmony-oriented results as needed. The experiment results show that the proposed unsupervised approach performs similar to or better than WordNet and NOMLEX. Furthermore, a new domain-verb association measure is presented to show the association relationships between inferred verbs and domains to which the verbs are possibly applied. 1 Introduction The variability of expression is an underlying phenomenon in natural language, and the recognition of the variability serves as the foundation of understanding natural language. Recognizing textual entailment is a research area that seeks to understand this variability, and thus to identify, generate, or extract textual entailment relations from texts. Textual entailment describes a relation of texts where the meaning of one text can be inferred plausibly from another text (Dagan et al., 2010). As a related term, paraphrases refer to expressions that deliver almost the same information using different words (Androutsopoulos and Malakasiotis, 2009). As a lighter form of textual entailment, inference rules refer to expressions that carry not only the same meanings but also similar meanings and could be useful to question answering (Lin and Pantel, 2001). Much research in recent years has focused on recognizing textual entailment pairs in natural language texts. For example, consider the following sentences: (1) Emily Bronte wrote Wuthering Heights. (2) Emily Bronte authored Wuthering Heights. Given that these two sentences deliver the same meaning, the verbs wrote and authored are in a textual entailment relation. Textual entailment plays a very important role in many areas. For example, in question answering, paraphrases from bilingual parallel corpus were used to expand the original questions (Lin and Pantel, 2001; Duboue and Chu-Carroll, 2006; Riezler et al., 2007); in information extraction, paraphrases were extracted and then used to find entities to fill the slots of binary relations (Shinyama and Sekine, 2003); in machine translation, paraphrases were captured and used as part of reference translations (Madnani et al., 2007; Marton et al., 2009). This work is licenced under a Creative Commons Attribution 4.0 International …

Content truncated to 500 words