Domain Representative Keywords Selection: A Probabilistic Approach

Findings of the Association for Computational Linguistics: ACL 2022, pages 679 - 692 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Domain Representative Keywords Selection: A Probabilistic Approach Pritom Saha Akash1 Jie Huang1 Kevin Chen-Chuan Chang1 Yunyao Li2∗ Lucian Popa3 ChengXiang Zhai1 1University of Illinois at Urbana-Champaign, USA 2Apple, USA 3IBM Research, USA {pakash2, jeffhj, kcchang, czhai}@illinois.edu yunyaoli@apple.com, lpopa@us.ibm.com Abstract We propose a probabilistic approach to select a subset of a target domain representative key- words from a candidate set, contrasting with a context domain. Such a task is crucial for many downstream tasks in natural language process- ing. To contrast the target domain and the con- text domain, we adapt the two-component mix- ture model concept to generate a distribution of candidate keywords. It provides more im- portance to the distinctive keywords of the tar- get domain than common keywords contrasting with the context domain. To support the repre- sentativeness of the selected keywords towards the target domain, we introduce an optimiza- tion algorithm for selecting the subset from the generated candidate distribution. We have shown that the optimization algorithm can be efficiently implemented with a near-optimal approximation guarantee. Finally, extensive ex- periments on multiple domains demonstrate the superiority of our approach over other baselines for the tasks of keyword summary generation and trending keywords selection.1 1 Introduction Domain representative keywords are the core knowledge of a target domain of interest. A target domain can be a broad area of science like com- puter science (CS) or its sub-field artificial intelli- gence (AI). Acquiring domain representative key- words benefits various natural language processing (NLP) tasks such as information summarization, organization, and extraction. For instance, acquir- ing a set of domain representative keywords is an important first step in organizing domain knowl- edge with a taxonomy of keywords (Zhang et al., 2018). Moreover, tagging documents (Chen et al., 2017) with domain representative keywords helps ∗* This work was done while the author was at IBM Re- search, USA. 1Code and data are available at https://github.com/ pritomsaha/keyword-selection Figure 1: Screenshots of example applications. to facilitate search or recommendation in a domain. For another example, summarizing a domain us- ing its trending keywords for a specific time frame helps researchers get a snapshot of research trends or emerging areas of interest, e.g., new emerging security vulnerabilities. In reality, while representing a domain, the desired keywords often depend on a given context domain. E.g., if we are interested in representing the CS domain with the context of general knowl- edge (all areas of knowledge), keywords like model, data, information make sense in distinguishing CS from general knowledge. However, if the context is general science, those keywords are not distinctive enough to distinguish CS from other areas like mathematics or physics. Instead, the keywords machine learning, data mining, deep learning make more sense in this case. Therefore, it is important to contrast with a known context domain while representing a target domain, but most of the existing work ignored this. An application of this …

Content truncated to 500 words