Semantic relevant sequential keyphrase generation using deep reinforcement learning

Abstract

In today s data-driven world, extracting or generating keyphrases is crucial for newlineprocessing huge volumes of text in NLP tasks. Keyphrase Extraction (KPE) and Keyphrase Generation (KPG) support applications such as SEO, summarization, and dialogue systems. Text generation, powered by Pre-trained Language Models (PLMs), has achieved considerable success in summarization, machine translation, and goaloriented dialogue (GoD). However, PLMs confront challenges such as limited integration of structured external knowledge and catastrophic forgetting. Although models like KeyBERT employ BERT embeddings and cosine similarity for ranking, they often ignore spatial and semantic diversity. Moreover, conventional graph-based KPE methods depend heavily on position and struggle to capture topic centrality or document context. In addition, existing KPG approaches face challenges in differentiating between present and absent keyphrases, resulting in poor generation of semantically rich absent phrases. These limitations emphasize the requirement for more robust, knowledge-integrated models that capture both contextual relevance and semantic depth in dynamic environments. To address these problems, three novel methodologies are proposed, including a text generation model for a GoD system (GoD-BERT) by finetuning the BERT on several GoD sub-tasks; an unsupervised graph and web-based KPE (GWebPositionRank) for improving the KeyBERT-based method with graph-based WebPositionRank; and multi-agent based deep reinforcement learning (RL) for sequential KPG (MADeGen) for efficiently differentiating present and absent keyphrases for sequential decisionmaking in KPG. The GoD-BERT model is proposed initially for addressing the newlinecatastrophic forgetting and insufficient goal-context understanding challenges in PLMs. It improves PLM performance through goal-specific preprocessing, goal-knowledge newlinegraph construction, and BERT fine-tuning through Graph Neural Networks (GNNs) and newlineadapter modules.