Enhancing Tree-Based Keyphrase Extraction Technique Using Embedding Technique

The automatic keyphrase extraction techniques aim to extract high-quality keyphrases for document summarization and many other purposes, including indexing and search optimization’ content classification and categorization, topic modeling and analysis, text mining, data analysis, and so on. However, most of the existing techniques are often domain-specific or require domain knowledge. Again, some of these methods employ complex statistical approaches that require significant computational resources. Others depend on large training data sets that may not always be available. To address these challenges, this paper proposes a new keyphrase extraction technique, named ETeKET or Enhanced Tree-Based Keyphrase Extraction Technique, which is domain-independent, requires minimal statistical knowledge, and does not rely on training data. It employs a special variant of binary tree, called Keyphrase Extraction (KePhEx) tree, to extract final keyphrases from candidate keyphrases. Moreover, to assess the degree of cohesiveness of various nodes concerning the root, it also employs a measure, called the Cohesiveness Index (CI), which offers flexibility to the process. Since the generation of final keyphrases depends on the creation of candidate keyphrases, the proposed technique integrates the KeyBERT model to leverage contextual embeddings for effective and efficient candidate keyphrase extraction. This integration enhances adaptability across different text lengths, with dynamic parameter tuning for optimal performance in varied contexts(e.g., short, medium, and long texts). The effectiveness of ETeKET is evaluated using three benchmark datasets, and the experimental results show significant improvement over TeKET, with an F1-score increase of up to 2.61 for SemEval-10, 1.6 for SemEval-2017, and a noteworthy gain of 6.89 for Thesis100 (top 10 results), highlighting ETeKET's enhanced performance, especially in diverse text types.