UBC_UOS-TYPED: Regression for typed-similarity

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 132–137, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics UBC UOS-TYPED: Regression for Typed-similarity Eneko Agirre University of the Basque Country Donostia, 20018, Basque Country e.agirre@ehu.es Nikolaos Aletras University of Sheffield Sheffield, S1 4DP, UK n.aletras@dcs.shef.ac.uk Aitor Gonzalez-Agirre University of the Basque Country Donostia, 20018, Basque Country agonzalez278@ikasle.ehu.es German Rigau University of the Basque Country Donostia, 20018, Basque Country german.rigau@ehu.es Mark Stevenson University of Sheffield Sheffield, S1 4DP, UK m.stevenson@dcs.shef.ac.uk Abstract We approach the typed-similarity task using a range of heuristics that rely on information from the appropriate metadata fields for each type of similarity. In addition we train a linear regressor for each type of similarity. The re- sults indicate that the linear regression is key for good performance. Our best system was ranked third in the task. 1 Introduction The typed-similarity dataset comprises pairs of Cul- tural Heritage items from Europeana1, a single ac- cess point to digitised versions of books, paintings, films, museum objects and archival records from in- stitutions throughout Europe. Typically, the items comprise meta-data describing a cultural heritage item and, sometimes, a thumbnail of the item itself. Participating systems need to compute the similarity between items using the textual meta-data. In addi- tion to general similarity, the dataset includes spe- cific kinds of similarity, like similar author, similar time period, etc. We approach the problem using a range of sim- ilarity techniques for each similarity types, these make use of information contained in the relevant meta-data fields.In addition, we train a linear regres- sor for each type of similarity, using the training data provided by the organisers with the previously de- fined similarity measures as features. We begin by describing our basic system in Sec- tion 2, followed by the machine learning system in 1http://www.europeana.eu/ Section 3. The submissions are explained in Section 4. Section 5 presents our results. Finally, we draw our conclusions in Section 6. 2 Basic system The items in this task are taken from Europeana. They cannot be redistributed, so we used the urls and scripts provided by the organizers to extract the corresponding metadata. We analysed the text in the metadata, performing lemmatization, PoS tagging, named entity recognition and classification (NERC) and date detection using Stanford CoreNLP (Finkel et al., 2005; Toutanova et al., 2003). A preliminary score for each similarity type was then calculated as follows: • General: cosine similarity of TF.IDF vectors of tokens, taken from all fields. • Author: cosine similarity of TF.IDF vectors of dc:Creator field. • People involved, time period and location: cosine similarity of TF.IDF vectors of loca- tion/date/people entities recognized by NERC in all fields. • Events: cosine similarity of TF.IDF vectors of event verbs and nouns. A list of verbs and nouns possibly denoting events was derived us- ing the WordNet Morphosemantic Database2. • Subject and description: cosine similarity of TF.IDF vectors of respective fields. IDF values were …

Content truncated to 500 words