2021, EMNLP, Latent Hatred: A Benchmark for Understanding Implicit Hate Speech (Dataset)

Paper
Latent Hatred: A Benchmark for Understanding Implicit Hate Speech

Introduction
The paper introduces a new benchmark for understanding implicit hate speech, which is defined as hate speech that uses coded or indirect language. It presents a taxonomy of implicit hate and a large-scale dataset with fine-grained labels for each message. The dataset is designed to fill a gap in the literature by providing resources to detect and understand the diverse forms of implicit hate speech that have been previously underexplored​.

Main Problem
The main problem addressed is the detection and classification of implicit hate speech, which is more challenging to recognize than explicit hate speech due to its indirect nature. The authors aim to create a comprehensive dataset and benchmark to aid in developing models capable of identifying implicit hate speech.

Illustrative Example
Explicit Hate: Jws and Nggers destroy & prevent everthing they touch.
Implicit Hate: How is Mexico doing these days? People come here because you couldn't build it. Target was the Mexicans and it implies that the Mexicans are incompetent

Input
A sentence which contains implicit hate or it is a non-hate sentence.

Output
Differentiating hate from non-hate and, if identified as hate, categorizing it into one of six fine-grained categories of hate speech within the established taxonomy.

Motivation

Related works and their gaps

Contribution of this paper
A new dataset focused on implicit hate speech, which includes fine-grained labels and a taxonomy of implicit hate categories.
A set of benchmark tasks designed to evaluate models' ability to detect and classify implicit hate speech.
The establishment of a framework for studying implicit hate in a more structured and systematic way​

Proposed methods

Experiments
Models SVM and BERT

Implementation

Gaps this work
https://github.com/GT-SALT/implicit-hate