Deftor at SemEval-2016 Task 14: Taxonomy enrichment using definition vectors

Proceedings of SemEval-2016, pages 1342–1345, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics Deftor at SemEval-2016 Task 14: Taxonomy Enrichment using Definition Vectors Hristo Tanev Joint Research Centre, EC via Enrico Fermi 2749 Ispra, Italy hristo.tanev@jrc.ec.europa.eu Agata Rotondi Universita’ Ca’ Foscari Venezia, Italy rotondiagata@gmail.com Abstract In this paper we describe the participation of the Joint Research Centre, EC, in task 14 - Semantic Taxonomy Enrichment at SemEval 2016. The algorithm which we propose trans- forms each candidate definition into a term vector, where each dimension represents a term and its value is calculated by TF.IDF. We attach the candidate term as a hyponym to the WordNet synset with the most similar definition. The results we obtained are en- couraging, considering the simplicity of our approach. The obtained F measure is below the average, but above one of the baselines. 1 Introduction In this paper we describe the participation of the Joint Research Centre, EC, in task 14 - Semantic Taxonomy Enrichment at SemEval 2016. We par- ticpate for the first time in a similar task. We opted for a relatively simple method for searching of rel- evant synsets, which does not exploit any external dictionary or another semantic resource. We called our system Deftor (DEFinition vecTOR). Deftor is a system which represents the definitions (glosses) as lexical vectors and finds the most similar one for each new lemma. Automatic enrichment of taxonomies and knowl- edge bases is very important especially for rapidly changing domain. The taxonomy enrichment task is quite challenging, mostly because of the many pos- sibilities when attaching a new term to an existing taxonomy: First, a new word can be attached as a hy- ponym to different concepts, which describe it at dif- ferent levels of abstraction. For example, in Word- Net the hurricane is a hyponym of cyclone, which is a hyponym of windstorm, which itself is a hyponym of storm and the storm is a hyponym of atmospheric phenomenon. It is not always easy to decide where to attach a concept: in the above mentioned case the definition of storm and windstorm are not very different. In this case, it is also difficult to decide if a new concept should be merged with a similar concept from WordNet or it should be attached as a hyponym. Another problem are the multiple as- pects from which a concept can be perceived. For example, one can consider hurricane to be a natural disaster. It is also a weather condition or cause of death. All these considerations unfortunately make taxonomy enrichment task quite ambiguous and dif- ficult to tackle. In some cases, the right attachment of a new concept will be difficult also for a human expert. Although the task is quite complicated, as we have pointed out, we applied a simple approach which is based on comparison of the lexical content of the definitions of the new concepts and the Word- Net synsets. Our approach to the taxonomy enrichment task represents each synset from …

Content truncated to 500 words