[Paper] Beyond Semantics: An Evidential Reasoning-Aware Multi-View Learning Framework for Trustworthy Mental Health Prediction

Source: arXiv - 2605.05121v1

Overview

This paper tackles a pressing problem in AI‑driven mental‑health tools: how to make predictions that are not only accurate but also trustworthy when the input text is noisy, ambiguous, or comes from a different distribution than the training data. By marrying semantic embeddings from encoder‑only models (e.g., BERT) with higher‑level reasoning cues from decoder‑only models (e.g., GPT‑style generators) and wrapping the whole pipeline in an evidential‑learning framework, the authors deliver a multi‑view system that can quantify how sure it is about each prediction.

Key Contributions

• Multi‑view architecture that fuses semantic representations (encoder‑only) with reasoning‑oriented representations (decoder‑only) for mental‑health classification.
• Evidential learning layer based on Subjective Logic that explicitly models belief, disbelief, and uncertainty for each view, enabling calibrated confidence scores.
• Evidence‑based fusion strategy that automatically discounts noisy or contradictory evidence while amplifying complementary signals.
• Comprehensive evaluation on three public mental‑health datasets (Dreaddit, SDCNL, DepSeverity) showing state‑of‑the‑art accuracy and well‑calibrated uncertainty.
• Robustness & interpretability analyses (noise injection, case studies) that demonstrate the model’s resilience and its ability to surface human‑readable reasoning traces.

Methodology

1. Two parallel encoders

   • Semantic view: a standard encoder‑only transformer (e.g., RoBERTa) processes the raw text to produce contextual token embeddings.
   • Reasoning view: a decoder‑only model is prompted to generate a short “reasoning summary” of the input (e.g., “The user expresses hopelessness”). Its hidden state serves as a higher‑level, inference‑oriented feature.

2. Evidential heads

   • Each view feeds into a small evidential classifier that outputs evidence for each class rather than a raw softmax.
   • Using Dirichlet distribution theory, the evidence is transformed into belief mass, disbelief mass, and an uncertainty mass (the latter grows when evidence is scarce or conflicting).

3. Subjective‑Logic fusion

   • The belief and uncertainty from both views are combined via the discounting and cumulative fusion operators of Subjective Logic.
   • The resulting fused Dirichlet parameters yield a final class probability and a calibrated uncertainty estimate.

4. Training objective

   • A evidential loss (negative log‑likelihood of the Dirichlet) encourages the model to assign high belief to correct classes while keeping uncertainty low on clean data.
   • An auxiliary regularization term penalizes over‑confident predictions on perturbed inputs, nudging the system to be cautious when evidence is weak.

The pipeline is end‑to‑end differentiable, so developers can plug in any encoder/decoder pair and fine‑tune on their own mental‑health corpora.

Results & Findings

• Performance boost: The multi‑view evidential model outperforms single‑view baselines (pure BERT or GPT) by 3–6 % absolute accuracy.
• Uncertainty quality: ECE drops by ~30 % compared to vanilla softmax, meaning confidence scores align much better with actual correctness.
• Noise robustness: When random word swaps or synonym replacements are injected, the model’s accuracy degrades gracefully (≤ 4 % drop) while uncertainty spikes, signaling the degradation to downstream users.
• Interpretability: The reasoning view’s generated summaries often highlight key mental‑health cues (e.g., “feeling isolated”), and the evidential scores can be visualized to show which view contributed most to a decision.

Practical Implications

• Risk‑aware deployment: Apps that screen for depression or suicidal ideation can now surface a trust score alongside the prediction, allowing clinicians to triage only high‑confidence cases.
• Dynamic model gating: Developers can set uncertainty thresholds to trigger human review, fallback to a simpler rule‑based system, or request additional user input.
• Plug‑and‑play architecture: Because the framework treats encoder and decoder as interchangeable modules, existing LLM APIs (OpenAI, Anthropic) can be wrapped without retraining the whole model.
• Regulatory friendliness: Evidential outputs provide a mathematically grounded uncertainty estimate, which aligns with emerging AI‑risk standards (e.g., EU AI Act) that demand transparency about model confidence.
• Cross‑domain adaptability: The same multi‑view evidential pattern can be ported to other high‑stakes NLP tasks—fraud detection, medical triage, or safety‑critical dialog systems—where over‑confidence is a liability.

Limitations & Future Work

• Computational overhead: Running both encoder‑only and decoder‑only models doubles inference latency, which may be prohibitive for real‑time mobile apps.
• Reasoning prompt design: The quality of the reasoning view hinges on carefully crafted prompts; automatic prompt optimization remains an open challenge.
• Dataset bias: The three benchmark corpora are English‑centric and collected from social‑media platforms, limiting generalization to clinical notes or non‑English populations.
• Future directions suggested by the authors include: (1) distilling the dual‑view system into a single lightweight model, (2) extending the evidential fusion to more than two views (e.g., multimodal signals like voice or facial expression), and (3) exploring active‑learning loops where high‑uncertainty cases are sent to clinicians for annotation, continuously improving the evidence base.

Authors

• Yucheng Ruan
• Ling Huang
• Qika Lin
• Kai He
• Mengling Feng

Paper Information

• arXiv ID: 2605.05121v1
• Categories: cs.CL
• Published: May 6, 2026
• PDF: Download PDF