Desarrollo de un algoritmo para el análisis de sentimientos de textos en Kichwa en el ámbito ecuatoriano

Abstract

In the context of cultural and linguistic diversity in Ecuador, the Constitution officially recognizes Kichwa as a means to promote intercultural dialogue. However, this language has been the object of prejudice and stigmatization in the culture and identity of indigenous peoples. Therefore, the state has implemented actions to promote the use of the language, including bilingual educational programs and the creation of cultural materials. Given the primarily oral nature of this language, there is little textual data available, which presents challenges for the development of computational linguistic analysis algorithms. In this research, a dataset was generated in the Kichwa language, manually labeled with lexical valence criteria, to evaluate the emotional charge of the tokens contained in the dictionary between positive, negative, or neutral. This dataset allows performing a sentiment analysis of a new text entered by the user in a web prototype developed using Flask and Python. In the development methodology, data pre-processing uses Natural Language Processing (NLP) techniques and similarity metrics such as the Jaccard coefficient and Vector Cosine are applied to quantify the polarity of the text entered in Kichwa. To validate this system, an experimentation phase was carried out that allowed us to evaluate the performance of our tool against two other sentiment analysis tools built with pre-trained text models. The results show that, when analyzing polarities of Kichwa texts, the developed model achieved a maximum accuracy of 95% and an improvement of 6% and 18% compared to the ChatGPT and Bard models, respectively.