CFILT-CORE: Semantic Textual Similarity using Universal Networking Language

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 216–220, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics CFILT-CORE: Semantic Textual Similarity using Universal Networking Language Avishek Dan IIT Bombay Mumbai, India avishekdan@cse.iitb.ac.in Pushpak Bhattacharyya IIT Bombay Mumbai, India pb@cse.iitb.ac.in Abstract This paper describes the system that was sub- mitted in the *SEM 2013 Semantic Textual Similarity shared task. The task aims to find the similarity score between a pair of sen- tences. We describe a Universal Network- ing Language (UNL) based semantic extrac- tion system for measuring the semantic simi- larity. Our approach combines syntactic and word level similarity measures along with the UNL based semantic similarity measures for finding similarity scores between sentences. 1 Introduction Semantic Textual Similarity is the task of finding the degree of semantic equivalence between a pair of sentences. The core Semantic Textual Similar- ity shared task of *SEM 2013 (Agirre et al., 2013) is to generate a score in the range 0-5 for a pair of sentences depending on their semantic similar- ity. Textual similarity finds applications in infor- mation retrieval and it is closely related to textual entailment. Universal Networking Language (UNL) (Uchida, 1996) is an ideal mechanism for seman- tics representation. Our system first converts the sentences into a UNL graph representation and then matches the graphs to generate the semantic relat- edness score. Even though the goal is to judge sen- tences based on their semantic relatedness, our sys- tem incorporates some lexical and syntactic similar- ity measures to make the system robust in the face of data sparsity. Section 2 give a brief introduction to UNL. Sec- tion 3 decribes the English Enconverter developed Figure 1: UNL Graph for ’The boy chased the dog’ Figure 2: UNL Graph for ’The dog was chased by the boy’ by us. Section 4 discusses the various similarity measures used for the task. Section 5 mentions the corpus used for training and testing. Section 6 describes the method used to train the system and Section 7 presents the results obtained on the task datasets. 2 Universal Networking Language Universal Networking Language (UNL) is an inter- lingua that represents a sentence in a language inde- pendent, unambiguous form. The three main build- ing blocks of UNL are relations, universal words and attributes. UNL representations have a graphical structure with concepts being represented as nodes (universal words) and interactions between concepts being represented by edges (relations) between the nodes. Figure 1 shows the UNL graph correspond- 216 ing to the sentence ’The boy chased the dog.’ The conversion from a source language to UNL is called enconversion. The reverse process of generating a natural language sentence from UNL is called de- conversion. The enconversion process is markedly more difficult than the deconversion process due to the inherent ambiguity and idiosyncracy of natural language. UNL representation captures the semantics inde- pendent of the structure of the language. Figures 1 and 2 show …

Content truncated to 500 words