A Domain-Agnostic Neurosymbolic Approach for Big Social Data Analysis: Evaluating Mental Health Sentiment on Social Media during COVID-19

Why it matters

• Social media generates large-scale, rapidly evolving signals (about 12 billion tweets during COVID-19) that can inform mental health surveillance.
• Purely data-driven models struggle with emerging terms, including pandemic-specific slang, which limits near-real-time monitoring quality.
• Large language model baselines can require high computational cost (more than 6 to 8 hours to converge in reported experiments), reducing practical rapid adaptation.
• Public-health deployment needs both high accuracy and adaptability across depression, anxiety, and addiction signals.

What we did

• Proposed a domain-agnostic neurosymbolic framework that integrates Word2Vec with multiple knowledge bases, including DSM-5, DAO, UMLS, and DBpedia.
• Introduced SEDO to modulate tweet embeddings through a learned weight matrix formulated via a Sylvester equation.
• Trained classifiers on semantically filtered data (900M to 600M tweets) and evaluated binary tasks for depression, addiction, and anxiety.
• Achieved F1 scores above 92%, outperforming zero-shot LLM baselines (about 70% to 80% F1) while converging in 40 to 55 minutes.
• Validated robustness with triangulation and ablation studies, including a +5.03% F1 gain for depression after SEDO fine-tuning.

How it works

• Semantic Gap Management (B1): Train domain-specific LDA and Word2Vec models; dynamically update lexicons with knowledge bases and neologisms.
• Metadata Scoring (B2): Compute semantic mapping and proximity scores; normalize index scores as supervision signals.
• SEDO-based Infusion: Solve a Sylvester equation to learn a weight matrix that aligns tweet and knowledge-base embedding spaces.
• Adaptive Classification (B3): Train binary classifiers (for example, Balanced Random Forest) for depression, addiction, and anxiety.
• Validation: Perform triangulation and ablation studies to measure robustness and component-level gains.

The architecture (Figure 1) shows how knowledge infusion and embedding modulation interact across three stages.

Key contributions

• A multi-stage neurosymbolic architecture that combines shallow and semi-deep knowledge infusion for dynamic social media analysis.
• Empirical results above 92% F1, with consistent gains over zero-shot LLM baselines (about 89% to 93.6% versus 70% to 80%).
• Demonstrated efficiency: 40 to 55 minute convergence versus more than 6 to 8 hours for LLM baselines under similar settings.
• Triangulation and ablation studies showing measurable SEDO and knowledge-integration effects (for example, +5.03% F1 for depression after fine-tuning).

As shown in Table I, the neurosymbolic approach consistently exceeds LLM performance in F1 score.

Table II quantifies the degradation when SEDO is removed, underscoring its impact.