Local Word Vectors Guide Keyphrase Extraction

Word vector representation techniques, built on word-word co-occurrence statistics, often provide representations that decode the differences in meaning between various words. This significant fact is a powerful tool that can be exploited to a great deal of natural language processing tasks. In this work, we propose a simple and efficient unsupervised approach for keyphrase extraction, called Reference Vector Algorithm (RVA) which utilizes a local word vector representation by applying the GloVe method in the context of one scientific publication at a time. Then, the mean word vector (reference vector) of the article’s abstract guides the candidate keywords’ selection process, using the cosine similarity. The experimental results that emerged through a thorough evaluation process show that our method outperforms the state-of-the-art methods by providing high quality keyphrases in most cases, proposing in this way an additional mode for the exploitation of GloVe word vectors.