MADeGen: Multi-Agent based Deep Reinforcement Learning for Sequential Keyphrase Generation
- Title
- MADeGen: Multi-Agent based Deep Reinforcement Learning for Sequential Keyphrase Generation
- Creator
- Jose J.; Soundarabai P.B.
- Description
- Keyphrase generation is an essential tool in the field of natural language processing for information retrieval, document summarization, and text recommendation applications, extracting succinct and representative phrases from the text document. Traditional keyphrase extraction methods applied the supervised or unsupervised learning fail to capture the sequential keyphrase generation in a dynamic environment. The keyphrase generation approaches lack focus on explicitly discriminating the present and absent keyphrases, leading to the inadequate generation of semantically rich absent keyphrases. Hence, this work utilizes the potential benefits of reinforcement learning with the design of a distinguished reward function for present and absent keyphrases for sequential decision-making in the keyphrase generation. Thus, this work presents a novel keyphrase generation system, MADeGen, utilizing Multi- Agent Deep Reinforcement Learning (MADRL). In particular, a multi-agent reinforcement system collaboratively enables the generation of representative and coherent keyphrases by the evaluation metric-aware cooperative reward function analysis and adaptively training the agents. The proposed MADeGen incorporates two major phases, such as multi-agent modelling and actor critic-based policy optimization towards accurate keyphrase generation. In the first phase, the proposed approach designs two learning agents, including the extraction agent and generation agent, with the incorporation of a pre-trained language model. In the multi-agent system, the generation agent is the finetuned version of the extraction agent with the integration of the Wikipedia source. Secondly, the evaluation-aware adaptive reward function is designed to evaluate each agent's generated keyphrases with reference to ground-truth keyphrases. In subsequence, the cooperative reward analysis triggers the actor critic-based policy optimization for the generation agent in the multi-agent system to precisely generate the semantically relevant keyphrases with the assistance of an external web source. Experimental results on several benchmark datasets, such as Inspec, PubMed, and wiki20, illustrate the effectiveness of the proposed MADeGen compared to the existing keyphrase extraction models, yielding state-of-the-art performance in keyphrase extraction tasks. The proposed MADeGen proves its higher performance in the present as well as absent keyphrase extraction as 0.367 and 0.438 F1-score, respectively, while testing on the Inspec dataset. (2024), (Intelligent Network and Systems Society). All Rights Reserved.
- Source
- International Journal of Intelligent Engineering and Systems, Vol-17, No. 6, pp. 69-85.
- Date
- 2024-01-01
- Publisher
- Intelligent Network and Systems Society
- Subject
- Actor-critic; Adaptive reward; Extraction agent; Generation agent; Keyphrase generation; Multi-agent; Reinforcement learning
- Coverage
- Jose J., Christ Deemed to be University, Bangalore, India; Soundarabai P.B., Christ Deemed to be University, Bangalore, India
- Rights
- Restricted Access
- Relation
- ISSN: 2185310X
- Format
- Online
- Language
- English
- Type
- Article
Collection
Citation
Jose J.; Soundarabai P.B., “MADeGen: Multi-Agent based Deep Reinforcement Learning for Sequential Keyphrase Generation,” CHRIST (Deemed To Be University) Institutional Repository, accessed February 24, 2025, https://archives.christuniversity.in/items/show/13453.