A supervised approach to taxonomy extraction using word embeddings

A supervised approach to taxonomy extraction using word embeddings Rajdeep Sarkar1, John P. McCrae2, Paul Buitelaar2 1Indian Institute of Technology Kharagpur 2Insight Centre for Data Analytics, National University of Ireland Galway rajdeep.sarkar@iitkgp.ac.in, john@mccr.ae, paul.buitelaar@insight-centre.org Abstract Large collections of texts are commonly generated by large organizations and making sense of these collections of texts is a significant challenge. One method for handling this is to organize the concepts into a hierarchical structure such that similar concepts can be discovered and easily browsed. This approach was the subject of a recent evaluation campaign, TExEval, however the results of this task showed that none of the systems consistently outperformed a relatively simple baseline.In order to solve this issue, we propose a new method that uses supervised learning to combine multiple features with a support vector machine classifier including the baseline features. We show that this outperforms the baseline and thus provides a stronger method for identifying taxonomic relations than previous methods. Keywords: taxonomy extraction, word embeddings, supervised learning, terminology 1. Introduction When confronted with a large collection of texts a key challenge is to organize these texts into a structure that makes it easy to discover the information that is behind these texts. For example, it is common that large orga- nizations will collect many documents generated by their staff into a content management system, with only lim- ited organization such as tags. One of the most widely used methodologies for organizing text collections is that of the taxonomic structure, where a set of topics (defined by a set of terms) are organized into a tree structure and individual documents are then clustered under these top- ics. This method, as exemplified in bibliothecal classifi- cation schemes such as the Dewey Decimal Classification scheme, has been widely employed by libraries to organize large collections of texts for over one hundred years. Such schemes are still widely employed by organizations such as the ACM1 or the IEEE (IEEE, 2017) to organize contribu- tions to major conferences and other research contributions. The task of organizing topics into a hierarchy is called tax- onomy extraction and consists of two key steps. Firstly, in the term extraction step the set of terms that are relevant can be extracted from a collection of texts and secondly, in the taxonomy learning step these terms are grouped into a hier- archical structure. The first step is well explored and recent strong results have been shown on this task (Astrakhant- sev, 2016; Buitelaar et al., 2013) and as such we shall not focus on it in the course of this article. The second task is however much less well-explored and this is the focus of this article, and so we assume that the terms have al- ready been identified by an approach such as those outlines above. The task of taxonomy extraction is closely related to tasks such as hypernym detection (Hearst, 1992) or ontol- ogy learning (Buitelaar et al., 2005), in which a structured representation of concepts should be learned. However, …

Content truncated to 500 words