JU-Evora: A Graph Based Cross-Level Semantic Similarity Analysis using Discourse Information

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 375–379, Dublin, Ireland, August 23-24, 2014. JU-Evora: A Graph Based Cross-Level Semantic Similarity Analysis using Discourse Information Swarnendu Ghosh Dept. of Computer Science and Engi- neering Jadavpur University, Kolkata, India swarbir@gmail. com Nibaran Das Dept. of Computer Science and Engi- neering Jadavpur University, Kolkata, India ni- baran@ieee.org Teresa Gonçalves Dept. of Informática University of Évora, Évora, Portugal tcg@evora.pt Paulo Quaresma Dept. of Informática University of Évora, Évora, Portugal pq@evora.pt Abstract Text Analytics using semantic information is the latest trend of research due to its potential to represent better the texts content compared with the bag-of-words approaches. On the contrary, representation of semantics through graphs has several advantages over the tradi- tional representation of feature vector. There- fore, error tolerant graph matching techniques can be used for text comparison. Neverthe- less, not many methodologies exist in the lit- erature which expresses semantic representa- tions through graphs. The present system is designed to deal with cross level semantic similarity analysis as proposed in the SemEval-2014 : Semantic Evaluation, Inter- national Workshop on Semantic Evaluation, Dublin, Ireland. 1 Introduction Text Analytics has been the focus of much re- search work in the last years. State of the art ap- proaches typically represent documents as vec- tors (bag-of-words) and use a machine learning algorithm, such as k-NN or SVM, to create a model and to compare and classify new docu- ments. However, and in spite of being able to obtain good results, these approaches fail to rep- resent the semantic content of the documents, losing much information and limiting the tasks that can be implemented over the document rep- resentation structures. To overcome these short- comings some research has been done aiming to use and evaluate more complex knowledge rep- resentation structures. In this paper, a new ap- proach which integrates a deep linguistic analysis of the documents with graph-based classification algorithms and metrics has been proposed. 2 Overview of the Task This task provides an evaluation for semantic similarity across different sizes of text, which we refer to as lexical levels. Specifically, this task encompasses four semantic similarity compari- sons:  paragraph to sentence(P2S),  sentence to phrase(S2Ph),  phrase to word(Ph2W), and  word to sense(W2S). Task participants were provided with pairs of each comparison type and asked to rate the pair according to the semantic similarity of the small- er item to the larger item. As an example, given a sentence and a paragraph, a system would assess how similar is the meaning of the sentence to the meaning of the paragraph. Ideally, a high- similarity sentence would reflect overall meaning of the paragraph. The participants were expected to assign a score between [0,4] to each pairs of sentences, where 0 shows no similarity in con- cept while 4 shows complete similarity in con- cept. 3 Theoretical Concepts 3.1 Discourse Representation Structures Extraction and representation of the information conveyed by texts can be performed through several approaches, starting from …

Content truncated to 500 words