LongDocRank: graph-augmented large language models for unsupervised keyphrase extraction from long documents

Keyphrase extraction is a crucial NLP task that identifies key information from extensive text corpora, aiding in content summarization, search, and information retrieval. This paper presents LongDocRank, an advanced unsupervised keyphrase extraction framework that enhances large language models (LLMs) through a graph-based ranking mechanism. In this approach, LLMs first generate a comprehensive set of candidate keyphrases from long-form text, which are then structured into a graph where nodes represent candidate keyphrases, and edges are established based on their co-occurrence within the document. This graph-based representation enables a more contextually aware ranking of keyphrases by leveraging structural relationships. We evaluated LongDocRank across three state-of-the-art LLMs using two widely used long-document benchmark datasets, comparing its performance against twenty-two baseline models. Experimental results demonstrate that LongDocRank significantly improves keyphrase extraction by effectively capturing the semantic and structural relationships within long and complex documents in an unsupervised manner. Source code is published on GitHub (https://github.com/hd10-iupui/LongDocRank).