iKernels-Core: Tree Kernel Learning for Textual Similarity

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 53–58, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics iKernels-Core: Tree Kernel Learning for Textual Similarity Aliaksei Severyn1 and Massimo Nicosia1 and Alessandro Moschitti1,2 1University of Trento, DISI, 38123 Povo (TN), Italy {severyn,m.nicosia,moschitti}@disi.unitn.it 2Qatar Foundation, QCRI, Doha, Qatar {amoschitti}@qf.org.qa Abstract This paper describes the participation of iKer- nels system in the Semantic Textual Similar- ity (STS) shared task at *SEM 2013. Different from the majority of approaches, where a large number of pairwise similarity features are used to learn a regression model, our model directly encodes the input texts into syntac- tic/semantic structures. Our systems rely on tree kernels to automatically extract a rich set of syntactic patterns to learn a similarity score correlated with human judgements. We ex- periment with different structural representa- tions derived from constituency and depen- dency trees. While showing large improve- ments over the top results from the previous year task (STS-2012), our best system ranks 21st out of total 88 participated in the STS- 2013 task. Nevertheless, a slight refinement to our model makes it rank 4th. 1 Introduction Comparing textual data to establish the degree of se- mantic similarity is of key importance in many Nat- ural Language Processing (NLP) tasks ranging from document categorization to textual entailment and summarization. The key aspect of having an accu- rate STS framework is the design of features that can adequately represent various aspects of the similar- ity between texts, e.g. using lexical, syntactic and semantic similarity metrics. The majority of approaches to semantic textual similarity treat the input text pairs as feature vec- tors where each feature is a score corresponding to a certain type of similarity. This approach is concep- tually easy to implement and STS-2012 (Agirre et al., 2012) has shown that the best systems were built following this idea, i.e. a number of features encod- ing similarity of an input text pair were combined in a single scoring model, such as Linear Regression or Support Vector Regression (SVR). One potential limitation of using only similarity features to repre- sent a text pair is that of low representation power. The novelty of our approach is that we encode the input text pairs directly into structural objects, e.g. trees, and rely on the power of kernel learning to ex- tract relevant structures. This completely different from (Croce et al., ), where tree kernels where used to establish syntactic similarity and then plugged as similarity features. To link the documents in a pair we mark the nodes in the related structures with a special relational tag. In this way effective struc- tural relational patterns are implicitly encoded in the trees and can be automatically learned by the kernel- based machine learning methods. We build our sys- tems on top of the features used by two best systems from STS-2012 and combine them with the tree ker- nel models within the Support Vector …

Content truncated to 500 words