SFS-TUE: Compound Paraphrasing with a Language Model and Discriminative Reranking

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 148–152, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics SFS-TUE: Compound Paraphrasing with a Language Model and Discriminative Reranking Yannick Versley SfS / SFB 833 University of T¨ubingen versley@sfs.uni-tuebingen.de Abstract This paper presents an approach for gener- ating free paraphrases of compounds (task 4 at SemEval 2013) by decomposing the train- ing data into a collection of templates and fillers and recombining/scoring these based on a generative language model and discrimina- tive MaxEnt reranking. The system described in this paper achieved the highest score (with a very small margin) in the (default) isomorphic setting of the scorer, for which it was optimized, at a disadvantage to the non-isomorphic score. 1 Introduction Compounds are an interesting phenomenon in nat- ural language semantics as they normally realize a semantic relation (between head and modifier noun) that is both highly ambiguous as to the type of rela- tion and usually nonambiguous as to the concepts it relates (namely, those of the two nouns). Besides inventory-based approaches, where the relation is classified into a fixed number of relations, many researchers have argued that the full variabil- ity of the semantic relations inherent in compounds is best captured with paraphrases: Lauer (1995) pro- poses to use a preposition as a proxy for the meaning of a compound. Finin (1980) and later Nakov (2008) and others propose less restrictive schemes based on paraphrasing verbs. A previous SemEval task (task 9 in 2010; But- nariu et al., 2009). The most successsful approaches for this task such as Nulty and Costello (2010), Li et al. (2010), and Wubben (2010), or the subse- quent approach of Wijaya and Gianfortoni (2011), all make efficient use of both the training data and general evidence from WordNet or statistics derived from large corpora. The paper of Li et al. men- tions that solely inducing a global ranking of para- phrasing verbs from the training data (looking which verb is ranked higher in those cases where both were considered for the same compound) yielded higher scores than an unsupervised approach based on the semantic resources, underlining the need to combine training data and resources efficiently. SemEval 2013 task 4 The present task on pro- viding free paraphrases for noun compounds (Hen- drickx et al., 2013) uses a dataset collected from Me- chanical Turk workers asked to paraphrase a given compound (without context). Prepositional, verbal, and other paraphrases all occur in the data: (1) a. bar for wine b. bar that serves wine c. bar where wine is sold d. sweet vinegar made from wine In the examples, the words of the compound (wine bar and wine vinegar, respectively) are put in ital- ics, and other content words in the paraphrase are underlined. It is clear that certain paraphrases (X for Y) will be common across many compounds, whereas the ones containing more lexical material will differ even be- tween relatively similar …

Content truncated to 500 words