janardhan: Semantic Textual Similarity using Universal Networking Language graph matching

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 662–666, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics janardhan: Semantic Textual Similarity using Universal Networking Language graph matching Janardhan Singh IIT Bombay, Mumbai, India janardhan @cse.iitb.ac.in Arindam Bhattacharya IIT Bombay, Mumbai, India arindamb @cse.iitb.ac.in Pushpak Bhattacharyya IIT Bombay, Mumbai, India pb @cse.iitb.ac.in Abstract Sentences that are syntactically quite different can often have similar or same meaning. The SemEval 2012 task of Semantic Textual Sim- ilarity aims at finding the semantic similarity between two sentences. The semantic repre- sentation of Universal Networking Language (UNL), represents only the inherent meaning in a sentence without any syntactic details. Thus, comparing the UNL graphs of two sen- tences can give an insight into how semanti- cally similar the two sentences are. This paper presents the UNL graph matching method for the Semantic Textual Similarity(STS) task. 1 Introduction Universal Networking language (UNL) gives the semantic representation of sentences in a graphi- cal form. By comparing the similarity of these graphs, we inherently compare only the semantic content of the two sentences, rather than compar- ing the similarities in the syntax. Thus, the UNL graph matching strategy is a natural choice for the Semantic Textual Similarity(STS) task of SemEval 2012. UNL graphs are also used in textual en- tailment and interlingua based machine translation tasks. We use the UNL enconverter system at: http://www.cfilt.iitb.ac.in /UNL enco to generate the UNL graphs of the sentences. For the two graphs, generated from the two sentences, we give a similarity score by matching the two graphs. In the following sections we describe UNL matching strategy. section 2 describes the UNL sys- Figure 1: UNL graph for “John eats rice” tem and why this approach is useful, section 3 de- scribes the matching algorithm, section 4 describes the challenges faced in this approach, section 5 gives the results and finally section 6 gives the conclusion and the future scope. 2 Universal Networking Language The Universal Networking Language gives a graph- ical representation of the semantics of a text in the form of hypergraphs. The representation is at the semantic level which allows mapping of the simi- lar meaning sentences having different syntax to the same representation. To exemplify this point, con- sider the UNL graphs generated for the following sentences: Sentence 1: John ate rice. Sentence 2: Rice was eaten by John. The UNL graph generated from the system are given in figures 1 and 2 respectively. The UNL graph consists of three components: 662 Figure 2: UNL graph for “Rice was eaten by John” • Universal Words • Relations • Attributes 2.1 Universal Words The Universal Words (UWs) form the vocabulary of the Universal Networking Language. They form the nodes of the UNL graph. The words are normalized to their basic lemma, for example, eats becomes eat. The Universal Word is, usually, followed by a dis- ambiguating constraint list which is mainly used for disambiguating the sense of the Universal Word. For example, John (iof > person), …

Content truncated to 500 words