KeyGames: A Game Theoretic Approach to Automatic Keyphrase Extraction

Proceedings of the 28th International Conference on Computational Linguistics, pages 2037–2048 Barcelona, Spain (Online), December 8-13, 2020 2037 KeyGames: A Game Theoretic Approach to Automatic Keyphrase Extraction Arnav Saxena⇤, Mudit Mangal⇤and Goonjan Jain Department of Applied Mathematics, Delhi Technological University New Delhi, India {arnavsaxenadtu,muditmangaldtu}@gmail.com, goonjan jain@hotmail.com Abstract In this paper, we introduce two advancements in the automatic keyphrase extraction (AKE) space - KeyGames and pke+. KeyGames is an unsupervised AKE framework that employs the concept of evolutionary game theory and consistent labelling problem to ensure consistent classification of candidates into keyphrase and non-keyphrase. Pke+ is a python based pipeline built on top of the existing pke library to standardize various AKE steps, namely candidate extraction and evaluation, to ensure truly systematic and comparable performance analysis of AKE models. In the experiments section, we compare the performance of KeyGames across three publicly available datasets (Inspec 2001, SemEval 2010, DUC 2001) against the results quoted by the existing state-of-the-art models as well as their performance when reproduced using pke+. The results show that KeyGames outperforms most of the state-of-the-art systems while generalizing better on input documents with different domains and length. Further, pke+’s pre-processing brings out improvement in several other system’s quoted performance as well. 1 Introduction Automatic Keyphrase Extraction (AKE) is a task of identifying important words and phrases that best describe a given text document. AKE makes searching and indexing large digital collections of text feasible and finds use in a variety of Natural Language Processing (NLP) and Information Retrieval (IR) tasks ((Zhang et al., 2004), (Hulth and Megyesi, 2006), (Berend, 2011)). Owing to its widespread use, Keyphrase Extraction has emerged as a fundamental NLP task, improvements in which could lead to improvements in higher-level applications that build upon it (Danesh et al., 2015). Due to this importance, many approaches to keyphrase extraction have been proposed in the literature, majorly along two research lines: supervised and unsupervised (Hasan and Ng, 2014). However, the field faces two specific problems. Firstly, with state-of-the-art performance on keyphrase extraction being much lower than on the many core NLP tasks, there is still a fair amount of improvement possible in this space (Papagiannopoulou and Tsoumakas, 2020). Secondly, there is no effective way to compare and analyze various past AKE systems since most of them use different experimental setups. Different works have used different pre-processing and evaluation to demonstrate their performance. These discrepancies in the past studies make assessing the progress of the AKE space very challenging. To deal with the first issue, we recognised the need to approach this challenging problem from a fresh perspective. We thus propose a game-theoretic framework for AKE - KeyGames. Players, strategies, and payoffs form an integral part of any game. In KeyGames, the candidate keyphrases act as players having two strategies each: being a keyphrase and being a non-keyphrase. Each phrase plays a game with all the other phrases before deciding upon its strategy. Next, the definition of keyphrases is paraphrased, such that it paves the way for us …

Content truncated to 500 words