UFSCar's Team at ABSAPT 2022: Using Syntax, Semantics and Context for Solving the Tasks

With the growth of social networks and commercial sites, an increasing number of users have shared their opinions and life experiences on websites. Aspect-Base Sentiment Analysis (ABSA) is a field dedicated to studying opinions and feelings expressed in text format, aiming to detect the polarity of feelings and relate them to entities. To achieve this goal, this task can be divided into two sub-tasks: Aspect Term Extraction and Sentiment Orientation Extraction. The first task aims to identify the aspect within a text review, while the Sentiment Orientation Extraction aims to infer the polarity of a text review concerning a single aspect. For the first one, spaCy was used as an approach for data pre-processing, tokenization, and feature extraction. Then the aspect terms were identified and extracted. For the second one, the approach consisted of inserting the relevant portion of the text review into the GoEmotions model and taking the top-3 predictions. If the top-1 prediction score was higher than a threshold, we would assign that polarity to the respective aspect. Otherwise, we would assign the top-2 and top-3 predictions, if they have the same polarity, to the respective aspect. In both cases, good results were achieved, yielding third place in the ABSAPT 2022 competition. forum in English and then translated via the itranslate library. The paper analyzes the fine-tuning performance of the BERTimbau-base and BERTimbau-large models in the emotions classification task. It uses the Class Balanced Loss (CB) method to address the dataset imbalance issue, proven by the domination of the neutral class. Compared to the results provided by the authors of the GoEmotions dataset, this work achieved better performance, attributed to the robustness against imbalance. Based on it, we would extract the relevant text review as an input of the BERT model, already pre-trained with the GoEmotions dataset, classifying it into 27 emotions mapped to positive, negative, and neutral polarities.