HOPE: Hybrid Optimized Parallel Encoding for Depression Symptom Detection

This repository contains the official implementation of HOPE (Hybrid Optimized Parallel Encoding), a novel framework for detecting and ranking depression symptoms in social media text.

Methodology Overview

HOPE addresses the inherent challenges through a three-component architecture:

1. Supervised Symptom Extractor

Leverages DepRoBERTa to capture depressive patterns from labeled training data and generate initial relevance scores:

• Base Model: DepRoBERTa, a RoBERTa model pre-trained on mental health forum data
• Task: Multi-label sentence classification for depression symptom categories
• Output: Initial symptom-sentence relevance scores from supervised learning

2. Unsupervised Intrinsic Semantic Clustering

Employs few-shot-enhanced symptom-initialized K-means clustering with paired-stream BERT embeddings to extract intrinsic semantic relationships directly from unlabeled testing inputs:

• Paired-Stream Embeddings: Combines nomic-embed-text-v1.5 and modernbert-embed-base for robust semantic representations
• Symptom Initialization: Cluster centroids initialized using BDI-II symptom descriptions
• Few-Shot Enhancement: Incorporates few-shot learning to improve cluster quality
• Output: Unsupervised symptom-sentence similarity scores based on semantic clustering

3. Optimized Hybrid Semantic Fusion

Leverages the dual-view of supervised and unsupervised scores via a learnable MLP module:

• Learnable Fusion: MLP-based fusion mechanism that learns optimal combination of both views
• Complementary Integration: Exploits complementary strengths of supervised patterns and unsupervised semantics
• Output: Final unified relevance scores yielding more stable and task-agnostic predictions

Repository Structure

.
├── clustering/           # Unsupervised module: embedding, K-means, filtering
├── data/                 # Raw and processed data
├── data_processing/      # Scripts for data loading and preparation
├── extractor/            # Supervised module: model, training, inference
├── figures/              # Model diagrams and plots
├── utils/                # Utility scripts and constants
├── evaluate.py           # Script for evaluating model predictions
├── fusion.py             # Script for fusing supervised and unsupervised scores
├── requirements.txt      # Python dependencies
└── README.md             # This file