Multitask learning aplicada à classificação de utilidade de opiniões

Abstract

The study investigates the application of Multitask Learning (MTL) in opinion helpfulness classification, considering the interaction between polarity (positive/negative) and helpfulness (helpful/not helpful) as complementary tasks. The research aims to develop a more precise and generalizable model for sentiment analysis and content filtering. Initially, the paper reviews related works, such as the study conduded by (SOUSA, 2023), which applies machine learning to classify helpful opinions in Brazilian Portuguese, and the approach by (SATAPATHY; PARDESHI; CAMBRIA, 2022), which employs MTL and BERT embeddings for polarity and subjectivity analysis. The theoretical framework covers topics such as Natural Language Processing (NLP), multitask learning, and opinion helpfulness classification. The methodology involves the use of UTLCorpus, a dataset containing 2,881,589 reviews of Android applications and movies, automatically labeled for helpfulness and polarity. The study compares Single-Task Learning (STL) and MTL models, utilizing neural networks and BERT to classify opinions and with that, metrics such as 0.64 for opinion utility classification and 0.91 for polarity classification using BERT were achieved.