Compositional Distributional Semantics Models in Chunk-based Smoothed Tree Kernels

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 93–98, Dublin, Ireland, August 23-24 2014. Compositional Distributional Semantics Models in Chunk-based Smoothed Tree Kernels Nghia The Pham University of Trento thenghia.pham@unitn.it Lorenzo Ferrone University of Rome “Tor Vergata” lorenzo.ferrone@gmail.com Fabio Massimo Zanzotto University of Rome “Tor Vergata” fabio.massimo.zanzotto@uniroma2.it Abstract The field of compositional distributional semantics has proposed very interesting and reliable models for accounting the distributional meaning of simple phrases. These models however tend to disregard the syntactic structures when they are ap- plied to larger sentences. In this paper we propose the chunk-based smoothed tree kernels (CSTKs) as a way to exploit the syntactic structures as well as the reliabil- ity of these compositional models for sim- ple phrases. We experiment with the rec- ognizing textual entailment datasets. Our experiments show that our CSTKs per- form better than basic compositional dis- tributional semantic models (CDSMs) re- cursively applied at the sentence level, and also better than syntactic tree kernels. 1 Introduction A clear interaction between syntactic and semantic interpretations for sentences is important for many high-level NLP tasks, such as question-answering, textual entailment recognition, and semantic tex- tual similarity. Systems and models for these tasks often use classifiers or regressors that exploit con- volution kernels (Haussler, 1999) to model both interpretations. Convolution kernels are naturally defined on spaces where there exists a similarity function be- tween terminal nodes. This feature has been used to integrate distributional semantics within tree kernels. This class of kernels is often referred to as smoothed tree kernels (Mehdad et al., 2010; Croce et al., 2011), yet, these models only use distribu- tional vectors for words. Compositional distributional semantics models (CDSMs) on the other hand are functions map- ping text fragments to vectors (or higher-order ten- sors) which then provide a distributional meaning for simple phrases or sentences. Many CDSMs have been proposed for simple phrases like non- recursive noun phrases or verbal phrases (Mitchell and Lapata, 2008; Baroni and Zamparelli, 2010; Clark et al., 2008; Grefenstette and Sadrzadeh, 2011; Zanzotto et al., 2010). Non-recursive phrases are often referred to as chunks (Abney, 1996), and thus, CDSMs are good and reliable models for chunks. In this paper, we present the chunk-based smoothed tree kernels (CSTK) as a way to merge the two approaches: the smoothed tree kernels and the models for compositional distributional se- mantics. Our approach overcomes the limitation of the smoothed tree kernels which only use vec- tors for words by exploiting reliable CDSMs over chunks. CSTKs are defined over a chunk-based syntactic subtrees where terminal nodes are words or word sequences. We experimented with CSTKs on data from the recognizing textual entailment challenge (Dagan et al., 2006) and we compared our CSTKs with other standard tree kernels and standard recursive CDSMs. Experiments show that our CSTKs perform better than basic compo- sitional distributional semantic models (CDSMs) recursively applied at the sentence level and better than syntactic tree kernels. The rest of the paper is organized as follows. …

Content truncated to 500 words