KnCe2013-CORE:Semantic Text Similarity by use of Knowledge Bases

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 138–142, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics KnCe2013-CORE:Semantic Text Similarity by use of Knowledge Bases Hermann Ziak Know-Center GmbH Graz University of Technology Inffeldgasse 13/ 6. Stock 8010 Graz, Austria hziak@know-center.at Roman Kern Know-Center GmbH Graz University of Technology Inffeldgasse 13/ 6. Stock 8010 Graz, Austria rkern@know-center.at Abstract In this paper we describe KnCe2013-CORE, a system to compute the semantic similarity of two short text snippets. The system com- putes a number of features which are gath- ered from different knowledge bases, namely WordNet, Wikipedia and Wiktionary. The similarity scores derived from these features are then fed into several multilayer perceptron neuronal networks. Depending on the size of the text snippets different parameters for the neural networks are used. The final out- put of the neural networks is compared to hu- man judged data. In the evaluation our system performed sufficiently well for text snippets of equal length, but the performance dropped considerably once the pairs of text snippets differ in size. 1 Introduction The task of the semantic sentence similarity is to as- sign a score to a given pair of sentences. This score should reflect the degree by which the two sentences represent the same meaning. The semantic similar- ity of two sentences could then be used in a num- ber of different application scenarios, for example it could help to improve the performance of informa- tion retrieval systems. In the past, systems based on regression mod- els in combination with well chosen features have demonstrated good performance on this topic[4] [6]. Therefore we took this approach as a starting point to develop our semantic similarity system; addition- ally, we integrated a number of existing knowledge bases into our system. With it, trained with the data discussed in the task specification of last year[1], we participated in the shared task of SEM 2013. Additionally, to the similarity based on the fea- tures derived from the external knowledge bases, we employ a neural network to compute the final simi- larity score. The motivation to use a supervised ma- chine learning algorithm has been the observation that the semantic similarity is heavily influenced by the context of the human evaluator. A financial ex- pert for example would judge sentences with finan- cial topics different to non financial experts, if oc- curring numbers differ from each other. The remainder of the paper is organised as fol- lows: In Section 2 we described our system, the main features and the neuronal network to combine different feature sets. In Section 3 the calculation method of our feature values is discribed. In Sec- tion 4 we report the results of our system based on our experiments and the submitted results of the test data. In Section 5 and 6 we discuss the results and the outcome of our work. 2 System Overview 2.1 Processing Initially the system puts the …

Content truncated to 500 words