A Random Walk Approach to Selectional Preferences Based on Preference Ranking and Propagation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1169–1179, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A Random Walk Approach to Selectional Preferences Based on Preference Ranking and Propagation∗ Zhenhua Tian†, Hengheng Xiang, Ziqi Liu, Qinghua Zheng‡ Ministry of Education Key Lab for Intelligent Networks and Network Security Department of Computer Science and Technology Xi’an Jiaotong University Xi’an, Shaanxi 710049, China {zhhtian†,qhzheng‡}@mail.xjtu.edu.cn Abstract This paper presents an unsupervised ran- dom walk approach to alleviate data spar- sity for selectional preferences. Based on the measure of preferences between predi- cates and arguments, the model aggregates all the transitions from a given predicate to its nearby predicates, and propagates their argument preferences as the given predi- cate’s smoothed preferences. Experimen- tal results show that this approach out- performs several state-of-the-art method- s on the pseudo-disambiguation task, and it better correlates with human plausibility judgements. 1 Introduction Selectional preferences (SP) or selectional restric- tions capture the plausibility of predicates and their arguments for a given relation. Kaze and Fodor (1963) describe that predicates and their arguments have strict boolean restrictions, either satisfied or violated. Sentences are semantically anomalous and not consistent in reading if they violated the restrictions. Wilks (1973) argues that “rejecting utterances is just what humans do not. They try to understand them.” He further states s- electional restrictions as preferences between the predicates and arguments, where the violation can be less preferred, but not fatal. For instance, given the predicate word eat, word food is likely to be its object, iPhone is likely to be implausible for it, and tiger is less preferred but not curious. SP have been proven to help many natural lan- guage processing tasks that involve attachment de- ∗Partial of this work was done when the first author vis- iting at Language Technologies Institute of Carnegie Mellon University sponsored by the China Scholarship Council. cisions, such as semantic role labeling (Resnik, 1993; Gildea and Jurafsky, 2002), word sense dis- ambiguation (Resnik, 1997), human plausibility judgements (Spasi´c and Ananiadou, 2004), syn- tactic disambiguation (Toutanova et al., 2005), word compositionality (McCarthy et al., 2007), textual entailment (Pantel et al., 2007) and pro- noun resolution (Bergsma et al., 2008) etc. A direct approach to acquire SP is to extract triples (q, r, a) of predicates, relations, and argu- ments from a syntactically analyzed corpus, and then conduct maximum likelihood estimation (M- LE) on the data. However, this strategy is infea- sible for many plausible triples due to data spar- sity. For example, given the relation <verb-dobj- noun> in a corpus, we may see plausible triples: eat - {food, cake, apple, banana, candy...} But we may not see plausible and implausible triples such as: eat - {watermelon, ziti, escarole, iPhone...} Then how to use a smooth model to alleviate data sparsity for SP? Random walk models have been successful- ly applied to alleviate the data sparsity issue on collaborative filtering in recommender systems. Many online businesses, such as Netflix, Ama- zon.com, and …

Content truncated to 500 words