ADAPT at SemEval-2018 Task 9: Skip-Gram Word Embeddings for Unsupervised Hypernym Discovery in Specialised Corpora

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 924–927 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics ADAPT at SemEval-2018 Task 9: Skip-Gram Word Embeddings for Unsupervised Hypernym Discovery in Specialised Corpora Alfredo Maldonado ADAPT Centre Trinity College Dublin Ireland Filip Klubiˇcka ADAPT Centre Dublin Institute of Technology Ireland firstname.lastname@adaptcentre.ie Abstract This paper describes a simple but competitive unsupervised system for hypernym discovery. The system uses skip-gram word embeddings with negative sampling, trained on specialised corpora. Candidate hypernyms for an input word are predicted based on cosine similar- ity scores. Two sets of word embedding mod- els were trained separately on two specialised corpora: a medical corpus and a music indus- try corpus. Our system scored highest in the medical domain among the competing unsu- pervised systems but performed poorly on the music industry domain. Our approach does not depend on any external data other than raw specialised corpora. 1 Introduction The SemEval-2018 shared task on Hypernymy Discovery sought to study approaches for iden- tifying words that hold a hypernymic relation (Camacho-Collados et al., 2018). Two words have a hypernymic relation if one of the words belongs to a taxonomical class that is more general than that of the other word. For example, the word ve- hicle belongs to a more general taxonomical class than car does, as car is a type of vehicle. Hy- pernymy can be seen as an is-a relationship. Hy- pernymy has been studied from different angles in the natural language processing literature as it is related to the human cognitive ability of generali- sation. This shared task differs from recent taxonomy evaluation tasks (Bordea et al., 2015, 2016) by concentrating on Hypernym Discovery: the task of predicting (discovering) n hypernym candidates for a given input word, within the vocabulary of a specific domain (Espinosa-Anke et al., 2016). This shared task provided a general language do- main vocabulary and two specialised domain vo- cabularies in English: medical and music indus- try. For each vocabulary, a reference corpus was also supplied. In addition to these English vo- cabularies, general language domain vocabularies for Spanish and Italian were also provided. The ADAPT team focused on the two specialised do- main English subtasks by developing an unsuper- vised system that builds word embeddings from the supplied reference corpora for these domains. Word embeddings trained on large corpora have been shown to capture semantic relations be- tween words (Mikolov et al., 2013a,b), including hypernym-hyponym relations. The word embed- dings built and used by the system presented here exploit this property. Although these word em- beddings do not distinguish one semantic relation from another, we expect that true hypernyms will constitute a significant proportion of the predicted candidate hypernyms. Indeed, we show that for the medical domain subtask, our system beats the other unsupervised systems, although it still ranks behind the supervised systems. Even though unsupervised systems tend to rank behind supervised systems in NLP tasks in gen- eral, our motivation to focus on an …

Content truncated to 500 words