| Summary: | Identifying humor and offense may prove to be an arduous task even for humans. It is, however, even more challenging to translate it into a logical process that a machine can understand. This work pretends to develop machine learning models which will be implemented to achieve this task. On this track, this study will be based on the SemEval 2021 workshop, where the participants were challenged to identify and score both humor and offense texts, as well as detect controversial sentences (SemEval 2021 - Task 7 - Detecting and Rating Humor and Offense), encouraging the use of current state-of-the-art algorithmic techniques in Natural Language Processing. The objective is to identify and propose the most optimal setup to achieve the highest performance on Humor Detection and related tasks using a common dataset aggregating eight thousand sentences classified with their respective binary humor indicator and humor rating, along with binary controversial indicators and offense rating values. This document presents a solution for the presented tasks based on BERT (Bidirectional Encoder Representations from Transformers) which makes use of Transformers interpreting the sentences in both directions (bidirectional), which brings a much higher context perception into the model. It will compare the performance of three different BERT variants (BERTBASE, DistillBERT, and RoBERTa), each of them designed for better fit on different tasks used by industry and academia. Concluding that DistillBERT presented the most accurate results in the Humor Detection and Humor Rating tasks, while RoBERTa performed best in the controversial detection task. Finally, BERTBASE outperformed in the Offensiveness Ranking task.
|
|---|