BUT-TYPED: Using domain knowledge for computing typed similarity

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 119–123, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics BUT-TYPED: Using domain knowledge for computing typed similarity Lubomir Otrusina Brno University of Technology Faculty of Information Technology IT4Innovations Centre of Excellence Bozetechova 2, 612 66 Brno Czech Republic iotrusina@fit.vutbr.cz Pavel Smrz Brno University of Technology Faculty of Information Technology IT4Innovations Centre of Excellence Bozetechova 2, 612 66 Brno Czech Republic smrz@fit.vutbr.cz Abstract This paper deals with knowledge-based text processing which aims at an intuitive notion of textual similarity. Entities and relations rel- evant for a particular domain are identified and disambiguated by means of semi-supervised machine learning techniques and resulting an- notations are applied for computing typed- similarity of individual texts. The work described in this paper particularly shows effects of the mentioned processes in the context of the *SEM 2013 pilot task on typed-similarity, a part of the Semantic Tex- tual Similarity shared task. The goal is to evaluate the degree of semantic similarity be- tween semi-structured records. As the evalu- ation dataset has been taken from Europeana – a collection of records on European cultural heritage objects – we focus on computing a se- mantic distance on field author which has the highest potential to benefit from the domain knowledge. Specific features that are employed in our sys- tem BUT-TYPED are briefly introduced to- gether with a discussion on their efficient ac- quisition. Support Vector Regression is then used to combine the features and to provide a final similarity score. The system ranked third on the attribute author among 15 submitted runs in the typed-similarity task. 1 Introduction The goal of the pilot typed-similarity task lied in measuring a degree of semantic similarity between semi-structured records. The data came from the Europeana digital library1 collecting millions of records on paintings, books, films, and other mu- seum and archival objects that have been digitized throughout Europe. More than 2,000 cultural and scientific institutions across Europe have contributed to Europeana. There are many metadata fields at- tached to each item in the library, but only fields title, subject, description, creator, date and source were used in the task. Having this collection, it is natural to expect that domain knowledge on relevant cultural heritage en- tities and their inter-relations will help to measure semantic closeness between particular items. When focusing on similarities in a particular field (a se- mantic type) that clearly covers a domain-specific aspect (such as field author/creator in our case), the significance of the domain knowledge should be the highest. Intuitively, the semantic similarity among authors of two artworks corresponds to strengths of links that can be identified among the two (groups of) authors. As the gold standard for the task resulted from a Mechanical Turk experiment (Paolacci et al., 2010), it could be expected that close fields corre- spond to authors that are well known to represent the same style, worked in the same …

Content truncated to 500 words