Estudo comparativo de desempenho de modelos de Embedding de Sentenças para agrupamento de artigos

Abstract

In this work, the performance of Sentece Embedding models for the article clustering problem is analyzed. A Web Scraping system was developed and collected 58716 article metadata. A performance analysis system of Sentece Embedding models was developed based on clustering using the K-Means algorithm that is trained with the base of collected articles and calculating the metrics of the resulting clust ers and verifying the impact of the embeddings generated by each model. The analyzed models were Doc2Vet, InferSent, Sentence-BERT and Universal Sentence Encoder. In the defined settings, which vary from using 9 classes with all 58716 samples from the database to 2 classes with a limitation of 500 samples per class, the Universal Sentence Encoder showed better values of v-measure and adjusted Rand index.