Comparación de algoritmos de resumen de texto para el procesamiento de editoriales y noticias en español

Language is affected not only by grammatical rules but also by the context and socio-cultural differences. Therefore, automatic text summarization, an area of interest in natural language processing (NLP), faces challenges such as identifying essential fragments according to the context and establishing the type of text under analysis. Previous literature has described several automatic summarization methods; however, no studies so far have examined their effectiveness in specific contexts and Spanish texts. In this paper, we compare three automatic summarization algorithms using news articles and editorials in Spanish. The three algorithms are extractive methods that estimate the importance of a phrase or word based on similarity or word frequency metrics. A document database was built with 33 editorials and 27 news articles, and three summaries of each text were manually extracted employing the three algorithms. The algorithms were quantitatively compared using the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metric. We analyzed the algorithms’ potential to identify the main components of a text. In the case of editorials, the automatic summary should include a problem and the author’s opinion. Regarding news articles, the summary should describe the temporal and spatial characteristics of an event. In terms of word reduction percentage and accuracy, the method based on the similarity matrix produced the best results and can achieve a 70 % reduction in both cases (i.e., news and editorials). However, semantics and context should be incorporated into the algorithms to improve their performance in terms of accuracy and sensitivity.