Semantic Framework for Comparison Structures in Natural Language

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 993–1002, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Semantic Framework for Comparison Structures in Natural Language Omid Bakhshandeh University of Rochester omidb@cs.rochester.edu James F. Allen University of Rochester Institute for Human and Machine Cognition james@cs.rochester.edu Abstract Comparison is one of the most impor- tant phenomena in language for express- ing objective and subjective facts about various entities. Systems that can under- stand and reason over comparative struc- ture can play a major role in the applica- tions which require deeper understanding of language. In this paper we present a novel semantic framework for represent- ing the meaning of comparative structures in natural language, which models com- parisons as predicate-argument pairs inter- connected with semantic roles. Our frame- work supports not only adjectival, but also adverbial, nominal, and verbal compara- tives. With this paper, we provide a novel dataset of gold-standard comparison struc- tures annotated according to our semantic framework. 1 Introduction Representing the meaning of text has long been a focus in linguistics and deriving computational models of meaning has been pursued by various semantic tasks such as semantic parsing. Deep semantic parsing (as opposed to shallow seman- tic parsing, such as semantic role labeling) aims to map a sentence in natural language into its cor- responding formal meaning representation (Zelle and Mooney, 1996; Berant and Liang, 2014). There has been a renewed interest in deeper se- mantic representations of natural language (Ba- narescu et al., 2013) in NLP community. Open- domain semantic representations enable inference and reasoning, which is required for many lan- guage understanding tasks such as reading com- prehension tests and open-domain question an- swering. Comparison is a common way for ex- pressing differences in sentiment and other prop- erties towards some entity. Comparison can hap- pen in very simple structures such as ‘John is taller than Sam’, or more complicated construc- tions such as ‘The table is longer than the sofa is wide’. So far the computational semantics of com- paratives and how they affect the meaning of text has not been studied effectively. That is, the dif- ference between the existing semantic and syntac- tic representation of comparatives is not distinc- tive enough for enabling deeper understanding of a sentence. For instance, the general logical form representation of the sentence ‘John is taller than Susan’ using the Boxer system (Bos, 2008) is the following: named(x0, john, per) & named(x1, susan, nam) &than(taller(x0), x1) (1) The above meaning representation does not fully capture the underlying semantics of the ad- jective ‘tall’ and what it means to be ‘taller’. A hu- man reader can easily infer that actually the height of John is greater than the height of Susan. An- other example to consider is the sentence ‘John is tall’, which basically has the typical logical form tall(john) –which is a very superficial represen- tation for the meaning of the predicate ‘tall’. Like- wise, a human reader can infer that defining …

Content truncated to 500 words