Structural Representations for Learning Relations between Pairs of Texts

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1003–1013, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Structural Representations for Learning Relations between Pairs of Texts Simone Filice and Giovanni Da San Martino and Alessandro Moschitti ALT, Qatar Computing Research Institute, Hamad Bin Khalifa University {sfilice,gmartino,amoschitti}@qf.org.qa Abstract This paper studies the use of structural representations for learning relations be- tween pairs of short texts (e.g., sentences or paragraphs) of the kind: the second text answers to, or conveys exactly the same information of, or is implied by, the first text. Engineering effective features that can capture syntactic and semantic re- lations between the constituents compos- ing the target text pairs is rather complex. Thus, we define syntactic and semantic structures representing the text pairs and then apply graph and tree kernels to them for automatically engineering features in Support Vector Machines. We carry out an extensive comparative analysis of state- of-the-art models for this type of relational learning. Our findings allow for achiev- ing the highest accuracy in two differ- ent and important related tasks, i.e., Para- phrasing Identification and Textual Entail- ment Recognition. 1 Introduction Advanced NLP systems, e.g., IBM Watson system (Ferrucci et al., 2010), are the result of effective use of syntactic/semantic information along with relational learning (RL) methods. This research area is rather vast including, extraction of syntac- tic relations, e.g., (Nastase et al., 2013), predicate relations, e.g., Semantic Role Labeling (Carreras and M`arquez, 2005) or FrameNet parsing (Gildea and Jurafsky, 2002) and relation extraction be- tween named entities, e.g., (Mintz et al., 2009). Although extremely interesting, the above methods target relations only between text con- stituents whereas the final goal of an intelligent system would be to interpret the semantics of larger pieces of text, e.g., sentences or para- graphs. This line of research relates to three broad fields, namely, Question Answering (QA) (Voorhees and Tice, 1999), Paraphrasing Identifi- cation (PI) (Dolan et al., 2004) and Recognition of Textual Entailments (RTE) (Giampiccolo et al., 2007). More generally, RL from text can be denied as follows: given two text fragments, the main goal is to derive relations between them, e.g., ei- ther if the second fragment answers the question, or conveys exactly the same information or is im- plied by the first text fragment. For example, the following two sentences: - License revenue slid 21 percent, however, to $107.6 million. - License sales, a key measure of demand, fell 21 percent to $107.6 million. express exactly the same meaning, whereas the next one: - She was transferred again to Navy when the American Civil War began, 1861. implies: - The American Civil War started in 1861. Automatic learning a model for deriving the re- lations above is rather complex as any of the text constituents, e.g., License revenue, a key measure of demand, in the two sentences plays an important role. Therefore, a suitable approach should ex- ploit representations …

Content truncated to 500 words