Leveraging Sentiment for Offensive Text Classification

TL;DR

This work investigates whether incorporating predicted sentiment can improve offensive language detection on the OLID dataset. By predicting sentiment for each instance and prepending it to the text (PPS-OLID), the authors compare lexical features, RNNs, and transformer models, with DeBERTa establishing the strongest baseline. They find that sentiment augmentation yields a modest, yet consistent F1-score gain (approximately 0.28) over the baseline DeBERTa, while primarily helping NOT predictions. The study demonstrates sentiment as a valuable auxiliary signal for offensive language classification and suggests directions for scaling to larger datasets and applying transfer learning to enrich sentiment coverage.

Abstract

In this paper, we conduct experiment to analyze whether models can classify offensive texts better with the help of sentiment. We conduct this experiment on the SemEval 2019 task 6, OLID, dataset. First, we utilize pre-trained language models to predict the sentiment of each instance. Later we pick the model that achieved the best performance on the OLID test set, and train it on the augmented OLID set to analyze the performance. Results show that utilizing sentiment increases the overall performance of the model.

Figures (3)

• Figure 1: Distribution of PPS per OLID labels.
• Figure 2: Confusion Matrix of a) DeBERTa and b) DeBERTa+PPS depicting effect of sentiment on the DeBERTa model.
• Figure 3: Heatmap on the Affect of Each Sentiment on DeBERTa model before and after PPS-OLID.