The key highlights and insights from the content are:
The authors define a new task called Implicit Target Span Identification (iTSI) that requires models to identify token spans within content that target protected groups, even when the targets are not explicitly stated.
To support research in this area, the authors create a new dataset called Implicit-Target-Span (ITS) by leveraging a pooling-based annotation approach. This allows them to capture both implicit and explicit references to target groups across three existing hate speech datasets: IHC, DynaHate, and SBIC.
The ITS dataset contains 57k annotated samples with an average of 1.7 target spans per sample. The authors find that the ITS dataset has nearly 20 times more unique targets compared to the original datasets, highlighting the prevalence of implicit targets.
The authors establish a baseline model called TargetDetect that uses a sequence tagging approach with transformer-based encoders like BERT and RoBERTa to identify target spans in the ITS dataset.
Experiments show that the RoBERTa-Large encoder achieves the best performance, with an F1 score of up to 80.8% on the ITS test sets. The model also performs competitively on the PLEAD dataset, a publicly available hate speech dataset.
Error analysis reveals common failure modes of the baseline model, such as boundary errors, discrepancies in the number of predicted and ground truth spans, and challenges with obfuscated, implicit, and subtle target references.
Overall, this work introduces a novel task and dataset to advance research on identifying implicit targets in hate speech, and provides a strong baseline model to build upon for future work.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Nazanin Jafa... at arxiv.org 04-01-2024
https://arxiv.org/pdf/2403.19836.pdfDeeper Inquiries