Standardizing Distress Analysis: Emotion-Driven Distress Identification and Cause Extraction (DICE) in Multimodal Online Posts

Gopendra Vikram Singh, Soumitra Ghosh, Atul Verma, Chetna Painkra, Asif EkbalEMNLP 2023 · DOI

Introduction

Researchers from IIT Patna have introduced DICE (Distress Identification and Cause Extraction), a framework designed to advance social media moderation beyond simple classification. Unlike traditional models that merely flag content as "distressed" (encompassing hate speech, abuse, or offensive posts), DICE identifies the presence of distress and extracts the specific textual phrases—or "causal spans"—that justify the label. By integrating multimodal inputs and a zero-shot emotion-driven strategy, the model achieves superior accuracy and interpretability.

The Multimodal Challenge

Automated hate speech detection is a high-stakes task often hindered by two major flaws: a reliance on text-centric analysis and a lack of transparency. Social media is inherently multimodal; a caption may appear harmless in isolation but becomes offensive when paired with a specific image. Furthermore, simply labeling a post as "offensive" is insufficient for legal or moderation workflows. DICE addresses these issues by treating distress analysis as a joint task of identification and rationale extraction.

The DICE Framework

The authors propose a multi-task deep learning framework to navigate the nuances of multimodal data:

  • Multimodal Encoding: The system utilizes BERT for textual analysis and ResNet for visual processing. To mitigate the "shattered gradients" problem common in deep training, the researchers implemented residual weights to maintain spatial structure.
  • Inter-modal Fusion (IMF): Rather than simple concatenation, the IMF module employs cross-attention to create "text-infused visual features" and "visual-infused textual features," allowing the image and text to provide mutual context.
  • Zero-Shot Emotion Integration: The model aligns features with Ekman’s six basic emotions (Anger, Sadness, etc.) using a zero-shot strategy. This allows the system to understand the affective state of a post without requiring explicit emotion labels for every sample.
  • Advanced Loss Functions: Using Insightface (ArcFace) loss, the model establishes sharper decision boundaries, increasing robustness against ambiguous samples.

The DCaM Dataset

A core contribution of this research is the DCaM (Distress and Cause annotated Multimodal) dataset. Comprising 20,764 posts scraped from Twitter and Gab, it is the first dataset to provide both multimodal content and human-annotated "rationales" for distress labels.

Example of a distressed post from the DCaM dataset with highlighted causal spans indicating the reason for the classification.

Key Insights: The Context Gap and Interpretability

A primary finding of the study is the "Context Gap." Many posts are impossible to classify correctly without visual context. For instance, a benign caption like "You don't have to be lonely" can be transformed into a hateful message when paired with specific political or extremist imagery.

A multimodal post demonstrating the 'context gap' where benign text becomes distressed when paired with specific imagery and hashtags.

By forcing the model to extract causal spans, the researchers move away from "black box" AI toward a "show me" approach. This interpretability is vital for human-in-the-loop moderation systems. Additionally, ablation studies confirmed that the Inter-modal Fusion and zero-shot emotion loss are critical to the model's performance.

Key Results

The DICE framework established new benchmarks on the DCaM dataset:

  • Distress Identification: Achieved an F1-score of 87.71%, outperforming previous baselines by nearly 2%.
  • Cause Extraction: Improved the ROS (Ratcliff-Obershelp Similarity) score by 3%, indicating that the model's extracted rationales closely align with human judgment.
  • Human Evaluation: In qualitative reviews, DICE was rated higher than SpanBERT and BERT-HateXplain in terms of fluency, knowledge consistency, and informativeness.

Limitations and Future Work

The authors note that keyword-based scraping may introduce sampling bias toward specific "hate words." Furthermore, current models often struggle with non-standard symbols (e.g., using "$" for "s") and emoticons used in toxic subcultures. Future research will focus on integrating these neglected features to make models more resilient to "coded" hate speech.

Summary updated

BibTeX
@inproceedings{singh2023standardizing,
  author = {Gopendra Vikram Singh and Soumitra Ghosh and Atul Verma and Chetna Painkra and Asif Ekbal},
  title = {{Standardizing Distress Analysis: Emotion-Driven Distress Identification and Cause Extraction (DICE) in Multimodal Online Posts}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.emnlp-main.275},
}
Figure