GETALP System : Propagation of a Lesk Measure through an Ant Colony Algorithm

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 232–240, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics GETALP: Propagation of a Lesk Measure through an Ant Colony Algorithm Didier Schwab, Andon Tchechmedjiev, Jérôme Goulian, Mohammad Nasiruddin, Gilles Sérasset, Hervé Blanchon LIG-GETALP Univ. Grenoble Alpes http://getalp.imag.fr/WSD firstname.lastname@imag.fr Abstract This article presents the GETALP system for the participation to SemEval-2013 Task 12, based on an adaptation of the Lesk measure propagated through an Ant Colony Algorithm, that yielded good results on the corpus of Se- meval 2007 Task 7 (WordNet 2.1) as well as the trial data for Task 12 SemEval 2013 (Ba- belNet 1.0). We approach the parameter es- timation to our algorithm from two perspec- tives: edogenous estimation where we max- imised the sum the local Lesk scores; exoge- nous estimation where we maximised the F1 score on trial data. We proposed three runs of out system, exogenous estimation with Ba- belNet 1.1.1 synset id annotations, endoge- nous estimation with BabelNet 1.1.1 synset id annotations and endogenous estimation with WordNet 3.1 sense keys. A bug in our imple- mentation led to incorrect results and here, we present an amended version thereof. Our sys- tem arrived third on this task and a more fine grained analysis of our results reveals that the algorithms performs best on general domain texts with as little named entities as possible. The presence of many named entities leads the performance of the system to plummet greatly. 1 Introduction Out team is mainly interested in Word Sense Disam- biguation (WSD) based on semantic similarity mea- sures. This approach to WSD is based on a local algorithm and a global algorithm. The local algo- rithm corresponds to a semantic similarity measure (for example (Wu and Palmer, 1994), (Resnik, 1995) or (Lesk, 1986)), while the global algorithm propa- gates the values resulting from these measures at the level of a text, in order to disambiguate the words that compose it. For two years, now, our team has focussed on researching global algorithms. The lo- cal algorithm we use, a variant of the Lesk algo- rithm that we have evaluated with several global al- gorithms (Simulated Annealing (SA), Genetic Al- gorithms (GA) and Ant Colony Algorithms (ACA)) (Schwab et al., 2012; Schwab et al., 2013), has shown its robustness with WordNet 3.0. For the present campaign, we chose to work with an ant colony based global algorithms that has proven its efficiency (Schwab et al., 2012; Tchechmedjiev et al., 2012). Presently, for this SemEval 2013 Task 12 (Nav- igli et al., 2013), the objective is to disambiguate a set of target words (nouns) in a corpus of 13 texts in 5 Languages (English, French, German, Italian, Spanish) by providing, for each sense the appropri- ate sense labels. The evaluation of the answers is performed by comparing them to a gold standard annotation of the corpus in all 5 languages using three possible sense inventories and thus …

Content truncated to 500 words