UMDuluth-CS8761 at SemEval-2018 Task9: Hypernym Discovery using Hearst Patterns, Co-occurrence frequencies and Word Embeddings

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 914–918 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics UMDuluth-CS8761 at SemEval-2018 Task 9: Hypernym Discovery using Hearst Patterns, Co-occurrence frequencies and Word Embeddings Arshia Z. Hassan and Manikya S. Vallabhajosyula and Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812 USA {hassa418,valla045,tpederse}@d.umn.edu https://github.com/manikyaswathi/SemEval2018HypernymDiscovery Abstract Hypernym Discovery is the task of identify- ing potential hypernyms for a given term. A hypernym is a more generalized word that is super-ordinate to more specific words. This paper explores several approaches that rely on co-occurrence frequencies of word pairs, Hearst Patterns based on regular expressions, and word embeddings created from the UMBC corpus. Our system Babbage participated in Subtask 1A for English and placed 6th of 19 systems when identifying concept hypernyms, and 12th of 18 systems for entity hypernyms. 1 Introduction Hypernym-hyponym pairs exhibit an is-a relation- ship where a hypernym is a generalization of a hy- ponym. The objective of SemEval–2018 Task 9 (Camacho-Collados et al., 2018) is to generate a ranked list of hypernyms when given an input hy- ponym and a vocabulary of candidate hypernyms. For example, the input hyponym lemongrass could yield the hypernyms [grass, oil plant, herb], where herb would be the best candidate. This scenario is illustrated in Figure 1, where the three leaf nodes are hyponyms and the root is a hypernym. Figure 1: Hypernym-hyponym example Note that hypernym discovery is distinct from hypernym detection, where the problem is to de- tect if a hyponym-hypernym relationship exists between a given pair, such as lemongrass-grass. In our first module, we retrieve candidate hy- pernyms for an input term using a paragraph- length context-window and calculate their co- occurrence frequencies, which is later used for ranking the candidates. Our second module uses Hearst Patterns (Hearst, 1992) to extract hyponym-hypernym pairs and ranks candidate hy- pernyms based on co-occurrence frequency of the pairs. Our final module employs word- embeddings created using word2vec (Mikolov et al., 2013). This paper continues with a more detailed discussion of each module, and then a re- view of our results. 2 Implementation Babbage begins by pre-processing (2.2.1) the UMBC Corpus (2.1) and extracting candidate hy- pernyms using four different strategies (2.2.2). The first and second module calculates the co- occurrence frequencies between the input term and words in context using the pre-processed UMBC Corpus and the Hearst Pattern set ex- tracted from the UMBC Corpus. The third mod- ule uses the IS-A Hearst Pattern set extracted from UMBC Corpus to obtain hypernyms. The final module constructs a word embedding over the UMBC corpus and uses a distance measure to fetch candidate hypernyms for a given input term. 2.1 UMBC Corpus Our training corpus is the University of Mary- land, Baltimore County (UMBC) WebBase Cor- pus (Han et al., 2013). It contains 3 billion words from paragraphs obtained from more than 100 mil- lion web pages over various domains. We use the 28 GB tokenized version of …

Content truncated to 500 words