[Paper Review] Hierarchical Attention Network for Explainable Depression Detection on Twitter Aided by Metaphor Concept Mappings

Sooji Han, Rui Mao+1 more|arXiv (Cornell University)|Sep 15, 2022

Mental Health via WritingPsychology20 citations

TL;DR

This paper presents an explainable Hierarchical Attention Network (HAN) for depression detection on Twitter that also leverages metaphor concept mappings (MCMs) to improve performance and interpretability. It achieves state-of-the-art results on a public MDL dataset and provides attention-based explanations for predictions.

ABSTRACT

Automatic depression detection on Twitter can help individuals privately and conveniently understand their mental health status in the early stages before seeing mental health professionals. Most existing black-box-like deep learning methods for depression detection largely focused on improving classification performance. However, explaining model decisions is imperative in health research because decision-making can often be high-stakes and life-and-death. Reliable automatic diagnosis of mental health problems including depression should be supported by credible explanations justifying models' predictions. In this work, we propose a novel explainable model for depression detection on Twitter. It comprises a novel encoder combining hierarchical attention mechanisms and feed-forward neural networks. To support psycholinguistic studies, our model leverages metaphorical concept mappings as input. Thus, it not only detects depressed individuals, but also identifies features of such users' tweets and associated metaphor concept mappings.

Motivation & Objective

• Motivate automatic, private depression screening on social media to aid early intervention while demanding interpretable models for health contexts.
• Introduce an explainable encoder that uses context-level attention to identify important tweets and metaphor mappings (MCMs) for each user.
• Incorporate MCM features into the depression detection pipeline to improve performance and provide cognitive insights.
• Evaluate the model against strong baselines on the MDL depression dataset and analyze explainability via attention visualizations.

Proposed method

• Embed tweets and MCMs with BERT-base-uncased to obtain T0 and C0 embeddings.
• Use a novel two-stream hierarchical attention encoder (HAN) to repeatedly refine tweet and MCM representations across layers (v_i^t, T_i and v_i^c, C_i).
• Concatenate final tweet and MCM representations and pass through a sequence of feed-forward networks to produce class probabilities.
• Incorporate Metaphor Concept Mappings (MCMs) via MI (metaphor identification), MP (metaphor paraphrasing), and CG (concept mapping generation) to generate MCM features from tweets.
• Train with cross-entropy loss to predict depressed vs non-depressed status.
• Provide qualitative explainability by inspecting attention weights over tweets and MCMs.

Experimental results

Research questions

• RQ1 Can a context-level, attention-based encoder provide interpretable depression predictions from Twitter data?
• RQ2 Do metaphor concept mappings (MCMs) offer complementary information that improves depression detection performance and interpretability?
• RQ3 How does the HAN encoder compare to classical encoders (LSTM, GRU, Transformer-based variants) in accuracy and parameter efficiency?
• RQ4 What is the impact of including MCM features on model performance across different data splits and datasets?

Key findings

• The proposed HAN model achieves state-of-the-art performance on the MDL dataset, with average F1 of 0.972 and accuracy of 0.971 across five splits.
• HAN outperforms the strongest baseline (Zogan et al. 2021) by 6.0% in F1 and 7.0% in accuracy on average.
• The ablation study shows MCMs provide additional information, with HAN-MCM lagging HAN by 1.1% F1 on MDL and 1.9% on IMDL, indicating MCMs help when explicit depression cues are removed.
• HAN encoders with 2 layers yield the best F1 (on par or better than other configurations) while using fewer parameters than LSTM/GRU baselines.
• Attention visualizations demonstrate explainability by highlighting the most influential tweets and MCMs for predicted depression status.
• The model remains robust across five different randomly sampled datasets, suggesting stable performance.