IIIT-H: A Corpus-Driven Co-occurrence Based Probabilistic Model for Noun Compound Paraphrasing

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 153–157, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics IIITH: A Corpus-Driven Co-occurrence Based Probabilistic Model for Noun Compound Paraphrasing Nitesh Surtani, Arpita Batra, Urmi Ghosh and Soma Paul Language Technologies Research Centre IIIT Hyderabad Hyderabad, Andhra Pradesh-500032 {nitesh.surtaniug08, arpita.batra, urmi.ghosh}@students.iiit.ac.in, soma@iiit.ac.in Abstract This paper presents a system for automatically generating a set of plausible paraphrases for a given noun compound and rank them in de- creasing order of their usage represented by the confidence value provided by the human annotators. Our system implements a corpus- driven probabilistic co-occurrence based model for predicting the paraphrases, that uses a seed list of paraphrases extracted from cor- pus to predict other paraphrases based on their co-occurrences. The corpus study reveals that the prepositional paraphrases for the noun compounds are quite frequent and well cov- ered but the verb paraphrases, on the other hand, are scarce, revealing the unsuitability of the model for standalone corpus-driven ap- proach. Therefore, to predict other paraphras- es, we adopt a two-fold approach: (i) Prediction based on Verb-Verb co- occurrences, in case the seed paraphrases are greater than threshold; and (ii) Prediction based on Semantic Relation of NC, otherwise. The system achieves a comparabale score of 0.23 for the isomorphic system while main- taining a score of 0.26 for the non-isomorphic system. 1 Introduction Semeval 2013 Task 4 (Hendrickx et. al., 2013), “Free Paraphrases of Noun Compounds” is a pa- raphrase generation task that requires the system to generate multiple paraphrases for a given noun compound and rank them to the best approxima- tion of the human rankings, represented by the cor- responding confidence value. The task is an extension of Semeval 2010 Task 9 (Butnariu et al., 2010), where the participants were asked to rank the set of given paraphrases for each noun com- pound. Although the ranking task is quite distinct from the task of generating paraphrases, however, we have taken many insights from the systems de- veloped for the ranking task, and have reported them appropriately in our system description. This paper describes a system for generating a ranked set of paraphrases for a given NC. A pa- raphrase can be Prepositional, Verb or Verb + Pre- positional. Since the prepositional paraphrases are easily available in the corpus while the occurrences of verb or verb+prep paraphrases is scarce, the task of paraphrasing becomes significant in finding out a method for predicting reliable paraphrases with verbs for a given NC. Our system implements a model that is based on co-occurrences of the pa- raphrases and selects those paraphrases that have a higher probability of co-occurring with a set of extracted paraphrases which are referred to as Seed Paraphrases. Keeping the verb-paraphrase scarcity issue in mind, we develop a two-way model: (i) Model 1 is used when the seed paraphrases are considerable in number i.e., greater than the thre- shold value. In this case, other …

Content truncated to 500 words