Keyword Extraction Approach Based on Probabilistic-Entropy, Graph, and Neural Network Methods

Nowadays, methods of automatic keyword extraction are developed based on statistical and graph features of texts. The transfer of learning approaches allows one to use additional word features obtained from deep neural network models fitted to solve different tasks. The paper proposes an integrated approach to keyword extraction based on a classification model that aggregates results of probabilistic-entropy, graph methods, and word features extracted from a neural network for text title generation. To validate the method, a dataset of news texts was gathered, with keywords manually selected through crowdsourcing. For the proposed approach F1-measure weighted by classes accuracy of keyword extraction is 72%, which is approximately 5% better in comparison with the existing methods.