ASAP: Automatic Semantic Alignment for Phrases

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 104–108, Dublin, Ireland, August 23-24, 2014. ASAP: Automatic Semantic Alignment for Phrases Ana O. Alves CISUC - University of Coimbra and Polytechnic Institute of Coimbra Portugal ana@dei.uc.pt Adriana Ferrugento Mariana Lourenc¸o Filipe Rodrigues CISUC - University of Coimbra Portugal {aferr,mrlouren}@student.dei.uc.pt fmpr@dei.uc.pt Abstract In this paper we describe the ASAP sys- tem (Automatic Semantic Alignment for Phrases)1 which participated on the Task 1 at the SemEval-2014 contest (Marelli et al., 2014a). Our assumption is that STS (Semantic Text Similarity) follows a func- tion considering lexical, syntactic, seman- tic and distributional features. We demon- strate the learning process of this function without any deep preprocessing achieving an acceptable correlation. 1 Introduction Evaluation of compositional semantic models on full sentences through semantic relatedness and textual entailment, title of this task on SemEval, aims to collect systems and approaches able to predict the difference of meaning between phrases and sentences based on their included words (Baroni and Zamparelli, 2010; Grefenstette and Sadrzadeh, 2011; Mitchell and Lapata, 2010; Socher et al., 2012). Our contribution is in the use of complemen- tary features in order to learn the function STS, a part of this challenge. Rather than specifying rules, constraints and lexicons manually, we advo- cate a system for automatically acquiring linguis- tic knowledge using machine learning (ML) meth- ods. For this we apply some preprocessing tech- niques over the training set in order to find differ- ent types of features. Related to the semantic as- pect, we make use of known semantic relatedness and similarity measures on WordNet, in this case, applied to see the relatedness/similarity between phrases from sentences. This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ 1This work was supported by the Crowds project- PTDC/EIA-EIA/115014/2009 Considering the problem of modeling a text cor- pus to find short descriptions of documents, we aim an efficient processing of large collections while preserving the essential statistical relation- ships that are useful for, in this case, similarity judgment. Therefore we also apply topic model- ing in order to get topic distribution over each sen- tence set. These features are then used to feed an ensemble algorithm to learn the STS function. 2 Background 2.1 WordNet WordNet (Miller, 1995) is a computational lexicon of English created and maintained at Princeton University. It encodes concepts in terms of sets of synonyms (called synsets). A synset can be seen as a set of word senses all expressing the same meaning. Each word sense uniquely identifies a single synset. For instance, car#n#1 uses the notation followed by WordNet and subscript word#p#n where p denotes the part-of-speech tag and n the word’s sense identifier, respec- tively. In this case, the corresponding synset car#n#1, auto#n#1, automobile#n#1, machine#n#6, motorcar#n#1 is uniquely determined. As words are not always so ambigu- ous, a word w#p is said to be monosemous when it can convey only …

Content truncated to 500 words