Topics Identification in a Corpus based on Transformers

Abstract

Classifying a corpus of texts based on a set of classes using Transformers allows to build a model based on the words that contain the information, however, the model uses all the words in the training process and the order in which its found, but, which of these words are directly related to the topic of the texts? This work focuses on proposing a methodology that allows, using a multi-labeled corpus that contains the description of 1200 comics organized in 4 classes, to eliminate information that are not related to the topic, based on the identification of named entities and noun phrases, thereby generating a new corpus with which to train and validate a Transformer, using the macro accuracy measure as an evaluation measure, as a base case the macro accuracy value, obtained of the validation of a Transformer trained with the original data is proposed, demonstrating that when we using data related to the subject matter of the texts, the classification results are improved from 0.733 to 0.992