DERI&UPM: Pushing Corpus Based Relatedness to Similarity: Shared Task System Description

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 643–647, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics DERI&UPM: Pushing Corpus Based Relatedness to Similarity: Shared Task System Description Nitish Aggarwal∗ Kartik Asooja◦ Paul Buitelaar∗ ∗Unit for Natural Language Processing Digital Enterprise Research Institute National University of Ireland, Galway, Ireland firstname.lastname@deri.org ◦Ontology Engineering Group Universidad Politecnica de Madrid Madrid, Spain asooja@gmail.com Abstract In this paper, we describe our system submit- ted for the semantic textual similarity (STS) task at SemEval 2012. We implemented two approaches to calculate the degree of simi- larity between two sentences. First approach combines corpus-based semantic relatedness measure over the whole sentence with the knowledge-based semantic similarity scores obtained for the words falling under the same syntactic roles in both the sentences. We fed all these scores as features to machine learn- ing models to obtain a single score giving the degree of similarity of the sentences. Lin- ear Regression and Bagging models were used for this purpose. We used Explicit Semantic Analysis (ESA) as the corpus-based seman- tic relatedness measure. For the knowledge- based semantic similarity between words, a modified WordNet based Lin measure was used. Second approach uses a bipartite based method over the WordNet based Lin measure, without any modification. This paper shows a significant improvement in calculating the semantic similarity between sentences by the fusion of the knowledge-based similarity mea- sure and the corpus-based relatedness measure against corpus based measure taken alone. 1 Introduction Similarity between sentences is a central concept of text analysis, however previous studies about semantic similarities have mainly focused either on single word similarity or complete document similarity. Sentence similarity can be defined by the degree of semantic equivalence of two given sen- tences, where sentences are typically 10-20 words long. The role of sentence semantic similarity mea- sures in text-related research is increasing due to potential number of applications such as document summarization, question answering, information extraction & retrieval and machine translation. One plausible limitation of existing methods for sentence similarity is their adaptation from long text (e.g. documents) similarity methods, where word co-occurrence plays a significant role. However, sentences are too short, thats why taking syntac- tic role of each word with its narrow semantic meaning into account, can be highly relevant to reflect the semantic equivalence of two sentences. These narrow semantics can be reflected from any existing large lexicons [(Wu and Palmer, 1994) and (Lin, 1998)]; nevertheless, these lexicons can not provide the semantics of words which are out of lexicon (e.g. guy) or multiword expressions. These semantics can be represented by a large distributed semantic space such as Wikipedia and similarity can be reflected by relatedness of these extracted semantics. However, relatedness covers broader space than similarity, which forced us to tune the Wikipedia based relatedness with lexical structure (e.g. WordNet) based similarities driven by linguistic syntactic structure, in reflecting more sophisticated similarity of two given sentences. In this work, we present a sentence similarity using ESA and syntactic similarities. …

Content truncated to 500 words