A Unified Kernel Approach for Learning Typed Sentence Rewritings

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 939–949, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics A Unified Kernel Approach for Learning Typed Sentence Rewritings Martin Gleize LIMSI-CNRS, Orsay, France Universit´e Paris-Sud, Orsay, France gleize@limsi.fr Brigitte Grau LIMSI-CNRS, Orsay, France ENSIIE, Evry, France bg@limsi.fr Abstract Many high level natural language process- ing problems can be framed as determin- ing if two given sentences are a rewrit- ing of each other. In this paper, we pro- pose a class of kernel functions, referred to as type-enriched string rewriting ker- nels, which, used in kernel-based machine learning algorithms, allow to learn sen- tence rewritings. Unlike previous work, this method can be fed external lexical se- mantic relations to capture a wider class of rewriting rules. It also does not assume preliminary syntactic parsing but is still able to provide a unified framework to cap- ture syntactic structure and alignments be- tween the two sentences. We experiment on three different natural sentence rewrit- ing tasks and obtain state-of-the-art results for all of them. 1 Introduction Detecting implications of sense between state- ments stands as one of the most sought-after goals in computational linguistics. Several high level tasks look for either one-way rewriting between single sentences, like recognizing textual entail- ment (RTE) (Dagan et al., 2006), or two-way rewritings like paraphrase identification (Dolan et al., 2004) and semantic textual similarity (Agirre et al., 2012). In a similar fashion, selecting sen- tences containing the answer to a question can be seen as finding the best rewritings of the ques- tion among answer candidates. These problems are naturally framed as classification tasks, and as such most current solutions make use of super- vised machine learning. They have to tackle sev- eral challenges: picking an adequate language rep- resentation, aligning semantically equivalent el- ements and extracting relevant features to learn the final decision. Bag-of-words and by extension bag-of-ngrams are traditionally the most direct ap- proach and features rely mostly on lexical match- ing (Wan et al., 2006; Lintean and Rus, 2011; Jimenez et al., 2013). Moreover, a good solving method has to account for typically scarce labeled training data, by enriching its model with lexical semantic resources like WordNet (Miller, 1995) to bridge gaps between surface forms (Mihalcea et al., 2006; Islam and Inkpen, 2009; Yih et al., 2013). Models based on syntactic trees remain the typical choice to account for the structure of the sentences (Heilman and Smith, 2010; Wang and Manning, 2010; Socher et al., 2011; Calvo et al., 2014). Usually the best systems manage to com- bine effectively different methods, like Madnani et al.’s meta-classifier with machine translation met- rics (Madnani et al., 2012). A few methods (Zanzotto et al., 2007; Zanzotto et al., 2010; Bu et al., 2012) use kernel func- tions to learn what makes two sentence pairs sim- ilar. Building on this work, we present a type- enriched string rewriting kernel giving the …

Content truncated to 500 words