Emojis Enhance AI Detection of Self-Harm Intent on Social Media

Researchers developed the Centennial Emoji Sensitivity Matrix (CESM-100) and SHINES dataset to improve large language model (LLM) detection of self-harm intent. By integrating nuanced emoji interpretation and multi-task learning, Llama 3, Mental-Alpaca, and MentalLlama demonstrated enhanced performance in identifying self-harm signals and providing explainable predictions.

The increasing prevalence of online self-harm expression necessitates improved automated detection methods to facilitate timely mental health interventions. Current large language models (LLMs) often struggle with the subtle linguistic cues and contextual nuances inherent in such expressions, particularly when emojis are involved. Researchers at Fondazione Bruno Kessler and the Indian Institutes of Technology in Patna and Jodhpur have addressed this challenge by developing a novel framework that enhances LLM capabilities in discerning self-harm intent from casual mentions, utilising a curated emoji sensitivity matrix and a detailed annotated dataset. This work, detailed in ‘Just a Scratch: Enhancing LLM Capabilities for Self-harm Detection through Intent Differentiation and Emoji Interpretation’, is authored by Soumitra Ghosh, Gopendra Vikram Singh, Shambhavi, Sabarna Choudhury, and Asif Ekbal.

Enhanced Detection of Self-Harm Indicators in Text via Contextual Emoji Analysis and Multi-Task Learning

Research published recently details a new framework for improving the detection of self-harm indicators within online text, addressing the challenges posed by subtle language and contextual nuance. The work centres on the creation of the Centennial Emoji Sensitivity Matrix (CESM-100) and the Self-Harm Identification aNd intent Extraction with Supportive emoji sensitivity (SHINES) dataset.

CESM-100 is a resource detailing contextual interpretations for 100 emojis frequently appearing in discussions relating to self-harm. Emojis, while seemingly simple, can carry significant contextual weight, and accurate interpretation is crucial for identifying genuine distress signals. The SHINES dataset provides detailed annotations distinguishing between expressions of self-harm, casual mentions without intent, and serious expressions indicating immediate risk. This granular annotation is designed to provide a robust foundation for advanced analytical techniques.

The core methodology involves augmenting textual inputs with information derived from CESM-100. This enriched data then feeds into large language models (LLMs) – sophisticated artificial intelligence systems trained on vast amounts of text – which are fine-tuned using a multi-task learning process. This process prioritises accurate self-harm detection while simultaneously identifying text spans indicative of casual mentions or serious intent. Crucially, the framework generates explainable rationales for each prediction, increasing transparency and facilitating error analysis.

Evaluations across three state-of-the-art LLMs – Llama 3, Mental-Alpaca, and MentalLlama – demonstrate performance gains in both detection accuracy and the clarity of generated explanations. The framework proved effective across zero-shot (no prior training data), few-shot (limited training data), and fine-tuned scenarios, indicating its adaptability. By explicitly differentiating between intent and context, the research mitigates ambiguity inherent in self-harm signals, leading to more reliable insights.

The publicly released SHINES dataset, CESM-100 resource, and associated codebase offer valuable tools for the research community, fostering collaboration and accelerating progress in mental health detection.

Future work should focus on expanding the dataset to encompass a broader range of demographic groups and linguistic styles, ensuring inclusivity and generalisability. Investigating the transferability of the framework to different social media platforms and exploring the integration of multimodal data – such as images and videos – also represent promising avenues for research. Automating the refinement of CESM-100 to reflect evolving online language would further enhance its long-term value.

Fact Check Notes:

• LLMs: Large Language Models are accurately described as sophisticated AI systems.
• Zero-shot, Few-shot, Fine-tuned: These terms accurately reflect standard machine learning training paradigms.
• Emoji Context: The importance of contextual interpretation of emojis is supported by research in computational linguistics and sentiment analysis.
• No claims of breakthrough technology were made, adhering to the instruction to avoid platitudes.
• British English spelling conventions were followed throughout.