DeepPurple: Lexical, String and Affective Feature Fusion for Sentence-Level Semantic Similarity Estimation

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 103–108, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics DeepPurple: Lexical, String and Affective Feature Fusion for Sentence-Level Semantic Similarity Estimation Nikolaos Malandrakis1, Elias Iosif2, Vassiliki Prokopi2, Alexandros Potamianos2, Shrikanth Narayanan1 1Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, CA 90089, USA 2Department of ECE, Technical University of Crete, 73100 Chania, Greece malandra@usc.edu, iosife@telecom.tuc.gr, vprokopi@isc.tuc.gr, potam@telecom.tuc.gr, shri@sipi.usc.edu Abstract This paper describes our submission for the *SEM shared task of Semantic Textual Sim- ilarity. We estimate the semantic similarity between two sentences using regression mod- els with features: 1) n-gram hit rates (lexical matches) between sentences, 2) lexical seman- tic similarity between non-matching words, 3) string similarity metrics, 4) affective content similarity and 5) sentence length. Domain adaptation is applied in the form of indepen- dent models and a model selection strategy achieving a mean correlation of 0.47. 1 Introduction Text semantic similarity estimation has been an ac- tive research area, thanks to a variety of potential ap- plications and the wide availability of data afforded by the world wide web. Semantic textual similar- ity (STS) estimates can be used for information ex- traction (Szpektor and Dagan, 2008), question an- swering (Harabagiu and Hickl, 2006) and machine translation (Mirkin et al., 2009). Term-level simi- larity has been successfully applied to problems like grammar induction (Meng and Siu, 2002) and affec- tive text categorization (Malandrakis et al., 2011). In this work, we built on previous research and our sub- mission to SemEval’2012 (Malandrakis et al., 2012) to create a sentence-level STS model for the shared task of *SEM 2013 (Agirre et al., 2013). Semantic similarity between words has been well researched, with a variety of knowledge-based (Miller, 1990; Budanitsky and Hirst, 2006) and corpus-based (Baroni and Lenci, 2010; Iosif and Potamianos, 2010) metrics proposed. Moving to sentences increases the complexity exponentially and as a result has led to measurements of simi- larity at various levels: lexical (Malakasiotis and Androutsopoulos, 2007), syntactic (Malakasiotis, 2009; Zanzotto et al., 2009), and semantic (Rinaldi et al., 2003; Bos and Markert, 2005). Machine trans- lation evaluation metrics can be used to estimate lex- ical level similarity (Finch et al., 2005; Perez and Alfonseca, 2005), including BLEU (Papineni et al., 2002), a metric using word n-gram hit rates. The pi- lot task of sentence STS in SemEval 2012 (Agirre et al., 2012) showed a similar trend towards multi-level similarity, with the top performing systems utilizing large amounts of partial similarity metrics and do- main adaptation (the use of separate models for each input domain) (B¨ar et al., 2012; ˇSari´c et al., 2012). Our approach is originally motivated by BLEU and primarily utilizes “hard” and “soft” n-gram hit rates to estimate similarity. Compared to last year, we utilize different alignment strategies (to decide which n-grams should be compared with which). We also include string similarities (at the token and character level) and similarity of affective …

Content truncated to 500 words