Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bi-Attention HateXplain : Taking into account the sequential aspect of data during explainability in a multi-task context

Ghislain Dorian Tchuente Mondjo

TL;DR

The paper addresses the instability of attention-based explanations in HateXplain-style multitask hate-speech models. It proposes BiAtt-BiRNN-HateXplain, a multitask framework that uses a BiRNN-based attention learning layer to capture sequential dependencies and align explanations with predictions. On the HateXplain dataset, the approach improves explainability fidelity, reduces bias, and maintains competitive performance, demonstrating that sequence-aware, interpretable architectures can outperform purely post-hoc explanations. The work suggests that transparent, efficient architectures can enhance trust and reliability in hate-speech detection systems, with potential integration into larger models like BERT or LIME in future work.

Abstract

Technological advances in the Internet and online social networks have brought many benefits to humanity. At the same time, this growth has led to an increase in hate speech, the main global threat. To improve the reliability of black-box models used for hate speech detection, post-hoc approaches such as LIME, SHAP, and LRP provide the explanation after training the classification model. In contrast, multi-task approaches based on the HateXplain benchmark learn to explain and classify simultaneously. However, results from HateXplain-based algorithms show that predicted attention varies considerably when it should be constant. This attention variability can lead to inconsistent interpretations, instability of predictions, and learning difficulties. To solve this problem, we propose the BiAtt-BiRNN-HateXplain (Bidirectional Attention BiRNN HateXplain) model which is easier to explain compared to LLMs which are more complex in view of the need for transparency, and will take into account the sequential aspect of the input data during explainability thanks to a BiRNN layer. Thus, if the explanation is correctly estimated, thanks to multi-task learning (explainability and classification task), the model could classify better and commit fewer unintentional bias errors related to communities. The experimental results on HateXplain data show a clear improvement in detection performance, explainability and a reduction in unintentional bias.

Bi-Attention HateXplain : Taking into account the sequential aspect of data during explainability in a multi-task context

TL;DR

The paper addresses the instability of attention-based explanations in HateXplain-style multitask hate-speech models. It proposes BiAtt-BiRNN-HateXplain, a multitask framework that uses a BiRNN-based attention learning layer to capture sequential dependencies and align explanations with predictions. On the HateXplain dataset, the approach improves explainability fidelity, reduces bias, and maintains competitive performance, demonstrating that sequence-aware, interpretable architectures can outperform purely post-hoc explanations. The work suggests that transparent, efficient architectures can enhance trust and reliability in hate-speech detection systems, with potential integration into larger models like BERT or LIME in future work.

Abstract

Technological advances in the Internet and online social networks have brought many benefits to humanity. At the same time, this growth has led to an increase in hate speech, the main global threat. To improve the reliability of black-box models used for hate speech detection, post-hoc approaches such as LIME, SHAP, and LRP provide the explanation after training the classification model. In contrast, multi-task approaches based on the HateXplain benchmark learn to explain and classify simultaneously. However, results from HateXplain-based algorithms show that predicted attention varies considerably when it should be constant. This attention variability can lead to inconsistent interpretations, instability of predictions, and learning difficulties. To solve this problem, we propose the BiAtt-BiRNN-HateXplain (Bidirectional Attention BiRNN HateXplain) model which is easier to explain compared to LLMs which are more complex in view of the need for transparency, and will take into account the sequential aspect of the input data during explainability thanks to a BiRNN layer. Thus, if the explanation is correctly estimated, thanks to multi-task learning (explainability and classification task), the model could classify better and commit fewer unintentional bias errors related to communities. The experimental results on HateXplain data show a clear improvement in detection performance, explainability and a reduction in unintentional bias.
Paper Structure (18 sections, 3 equations, 8 figures, 4 tables)

This paper contains 18 sections, 3 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Hate speech detection
  4. Explicabily of black box models
  5. Local Interpretable Model-Agnostic Explanations (LIME)
  6. SHapley Additive exPlanations (SHAP)
  7. Layer-wise Relevance Propagation (LRP)
  8. HateXplain Models based on multitask learning
  9. Methodology : Take account a sequential aspect of data on HateXplain model
  10. Bidirectional Recurrent Neural Network (BiRNN)
  11. BiAtt-HateXplain(Bidirectional Attention HateXplain)
  12. BiAtt-BiRNN-HateXplain(Bidirection BiRNN Hatexplain)
  13. Experimentations
  14. HateXplain
  15. Evaluation Metrics
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Flow of the LIME approach to explain the prediction of a classifier $f$ on an instance $x$.18
  • Figure 2: Study of the SHAP approach to explain the prediction of a classifier $f$ on an instance $x$.
  • Figure 3: Attention predicted by BiRNN-HateXplain HateXplain in yellow and ground truth attention in green
  • Figure 4: Representation of the general architecture of the model.
  • Figure 5: Representation of BiRNN.
  • ...and 3 more figures