A Short Survey on Taxonomy Learning from Text Corpora: Issues, Resources and Recent Advances

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1190–1203 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics A Short Survey on Taxonomy Learning from Text Corpora: Issues, Resources and Recent Advances Chengyu Wang, Xiaofeng He∗, Aoying Zhou Shanghai Key Laboratory of Trustworthy Computing, School of Computer Science and Software Engineering, East China Normal University chywang2013@gmail.com, {xfhe,ayzhou}@sei.ecnu.edu.cn Abstract A taxonomy is a semantic hierarchy, con- sisting of concepts linked by is-a relations. While a large number of taxonomies have been constructed from human-compiled resources (e.g., Wikipedia), learning tax- onomies from text corpora has received a growing interest and is essential for long- tailed and domain-specific knowledge ac- quisition. In this paper, we overview re- cent advances on taxonomy construction from free texts, reorganizing relevant sub- tasks into a complete framework. We also overview resources for evaluation and dis- cuss challenges for future research. 1 Introduction A taxonomy is a semantic hierarchy that organizes concepts by is-a relations, which exhibits the capa- bility of improving many NLP and IR tasks, such as query understanding (Hua et al., 2017), per- sonalized recommendation (Zhang et al., 2014), question answering (Yang et al., 2017), etc. also supports a variety of real-world applica-It tions, including information management (Nicker- son et al., 2013), e-commerce (Aanen et al., 2015) and biomedical systems (K¨ohler et al., 2014). With massive Web data available, a num- ber of taxonomies are constructed from human- compiled resources such as Wikipedia, Wikidata, etc (Suchanek et al., 2007; Ponzetto and Navigli, 2009; Flati et al., 2014; Mahdisoltani et al., 2015). But even large taxonomies may lack domain- specific and long-tailed knowledge. Recently, sev- eral methods have been developed to induce tax- onomies from text corpora (Wu et al., 2012; Yang, 2012; Luu et al., 2014). However, this task is far from being solved for three reasons: i) Text ∗Corresponding author. corpora may vary in size, topic and quality. is unlikely to develop a “one-size-fits-all” solu-It tion for all scenarios. For example, given an ex- tremely large corpus, Hearst-pattern based method is employed to build Probase (Wu et al., 2012). For domain-specific corpora, learning hypernymy embedding is more suitable (Luu et al., 2016b). ii) The accuracy of free-text taxonomies is usu- ally lower than many Wikipedia-based taxonomies because it is difficult to extract knowledge com- pletely from texts; iii) The task of taxonomy learn- ing is still insufficiently studied a) in emerging and specific domains and b) for non-English or under- resourced languages (Wei et al., 2014; Alfarone and Davis, 2015; Wang et al., 2015). In this paper, we overview recent advances on taxonomy construction from text corpora, reorga- nizing relevant subtasks into a complete frame- work. The subtasks include hyponym acquisi- tion, hypernym prediction, taxonomy induction, etc. We also summarize resources, evaluation met- rics and state-of-the-art results. We also discuss issues and directions for future research. 2 Taxonomy Construction Techniques Although workflows of different methods vary, a free text-based taxonomy construction system typ- ically operates in two steps: …

Content truncated to 500 words