🤖 machine learning

Beyond Accuracy: Interpreting Topic Representation in Suicide Ideation Detection Models

This paper demonstrates that topic-aware data augmentation not only enhances the performance of suicide ideation detection models but also improves the clarity, distinctness, and interpretability of their internal representations of critical psychosocial risk factors.

Original authors: Hamideh Ghanadian, Isar Nejadgholi, Hussein Al Osman

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hamideh Ghanadian, Isar Nejadgholi, Hussein Al Osman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart robot assistant trained to read social media posts and spot people who might be thinking about suicide. Usually, when we check if this robot is doing a good job, we just look at a score: "Did it get the right answer 90% of the time?"

But this paper asks a different question: "How does the robot actually think?"

The authors are worried that even if the robot gets the right answer, it might be doing so for the wrong reasons. Maybe it's just looking for the word "sad" and ignoring everything else. If the robot is a black box, we can't trust it in high-stakes situations like mental health.

Here is what they did, explained simply:

1. The Problem: The Robot's "Mental Map" is Messy

Think of the robot's brain as a giant, invisible map where it stores ideas.

The Issue: In the original training data, some topics (like "depression" or "anxiety") appear all the time. Other important topics (like "immigration struggles" or "family fights") are rare.
The Result: The robot's map gets cluttered. It mixes up different ideas. For example, it might not have a clear, separate "folder" for "financial crisis." Instead, it might mix "financial crisis" with "general sadness" or "anger." This is called entanglement. The robot is confused because it can't clearly separate one risk factor from another.

2. The Solution: Giving the Robot a "Topic Dictionary"

The researchers tried a new training method called Topic-Aware Augmentation.

The Analogy: Imagine you are teaching a child to identify fruits. If you only show them red apples, they might think "red" means "apple." But if you show them green apples, yellow bananas, and purple grapes, they learn that "fruit" is a big category with many distinct types.
What they did: They used AI to generate new, fake social media posts that specifically talked about rare topics (like immigration or money problems) and mixed them into the training data. This forced the robot to pay attention to these specific, rare risks.

3. How They Looked Inside the Robot's Brain

To see if this new training helped, they didn't just check the final score. They used two special tools to peek inside the robot's "mental map":

Tool A: The "Cluster" Visualizer (UMAP)
Imagine taking all the robot's ideas and plotting them on a piece of paper.
- Before: The dots representing "family issues" were scattered everywhere, mixed in with dots for "anger" and "sadness." It was a messy soup.
- After: The dots for "family issues" clumped together into a neat, tight group. They had their own clear neighborhood on the map.
- The Takeaway: The robot now has distinct, organized "folders" for different risks.
Tool B: The "Angle" Ruler (Cosine Distance)
Imagine two arrows pointing in different directions. If they point in the same direction, they are confused. If they point in opposite directions, they are clear.
- They measured how far apart the "idea of financial crisis" was from "general random text."
- The Result: In the new model, the arrow for "financial crisis" pointed in a much sharper, more unique direction compared to the old model. It was less likely to get mixed up with other things.

4. What They Found

Rare Topics Got Clearer: Topics that were previously hidden or messy (like immigration, family issues, and financial crisis) became very clear and distinct in the robot's brain.
Common Topics Stayed Strong: The robot didn't lose its ability to recognize common things like "depression."
The "Ceiling" Effect: Some topics, like "unemployment," were already so clear in the old robot that they couldn't get much clearer. It's like trying to make a perfect circle even more perfect; there's only so much room for improvement.
The "Overlap" Exception: One topic, "hopelessness," didn't get much clearer. The authors suggest this is because "hopelessness" is so deeply mixed up with "depression" and "anxiety" in real life that it's hard for the robot to separate them completely.

Summary

The paper argues that accuracy isn't enough. A model can be right by accident. By using this new training method, they didn't just make the robot smarter; they made its internal thinking process more organized.

Instead of a messy pile of mixed-up ideas, the robot now has a structured library where specific risks (like money trouble or family fights) have their own clear, separate shelves. This makes the model safer and easier to understand, because we can see exactly how it is identifying a risk, rather than just guessing that it got the right answer.

Technical Summary: Beyond Accuracy: Interpreting Topic Representation in Suicide Ideation Detection Models

Problem Statement

Suicide ideation detection models are typically evaluated using aggregate performance metrics (e.g., accuracy, F1-score), yet their internal representations of psychologically meaningful risk factors remain largely opaque. In high-stakes mental health applications, this lack of interpretability hinders the identification of spurious correlations, the auditing of learned representations, and the assurance of safe deployment.

Current research indicates that existing datasets often suffer from topic imbalance: common risk factors (e.g., depression, anxiety) are over-represented, while others (e.g., immigration, financial crisis) are under-represented or absent. This imbalance can lead classifiers to develop false causal relationships with over-represented concepts or become under-sensitive to rare topics. Consequently, models may fail to form clear, distinct internal representations for all psychological risk factors, potentially entangling them with unrelated features (polysemanticity) or failing to capture them entirely.

The paper addresses two primary research questions:

How are psychological risk factors represented within the internal activation space of language models trained for suicide ideation detection?
Does topic-aware synthetic data augmentation improve the clarity, separability, and specialization of these risk-factor representations?

Methodology

The study shifts focus from model outputs to the model's internal representations, specifically the intermediate activation vectors in the residual stream. The methodology employs a mechanistic interpretability approach grounded in the linear representation hypothesis and the superposition hypothesis.

1. Data and Models

Datasets: The analysis utilizes the University of Maryland (UMD) Reddit Suicidality Dataset as a baseline. A second dataset is constructed using topic-aware augmentation, where synthetic suicide-related posts are generated via Large Language Models (LLMs) and combined with subsets of the original UMD data to balance topic coverage.
Base Model: An ALBERT-base-V2 model is fine-tuned separately on the original UMD dataset and the augmented dataset. ALBERT is chosen for its compact architecture and stable residual-stream representations.

2. Dictionary Learning via Sparse Autoencoders (SAEs)

To uncover interpretable feature directions, the authors train overcomplete sparse autoencoders on the token-level residual-stream activations of the fine-tuned ALBERT models.

Architecture: The SAE uses an overcomplete dictionary where the number of latent features ( $F=4096$ ) significantly exceeds the input dimension ( $D=768$ ). This design aims to disentangle features that are superimposed in the dense model activations.
Mechanism: The SAE decomposes activation vectors into sparse linear combinations of learned basis directions (decoder vectors). Each direction represents a potential feature, and the encoder output indicates the activation strength of that feature for a given token.
Feature-Topic Mapping: Features are assigned a "dominant topic" based on the average activation of the feature across posts labeled with specific psychological topics (e.g., "Family Issues," "Immigration").

3. Analysis Techniques

Two complementary methods are used to assess the structure and interpretability of the learned features:

UMAP Visualization: Uniform Manifold Approximation and Projection is applied to the normalized decoder vectors to visualize the geometric organization of feature directions in 2D. This allows for the qualitative assessment of whether features associated with the same topic form coherent, compact clusters (indicating monosemanticity) or are dispersed (indicating polysemanticity).
Cosine Distance for Separability: A quantitative metric is introduced to measure the angular separation between the mean activation vectors of a concept set (posts with a specific topic) and a non-concept baseline (unrelated Reddit posts).
- Token-level activations are aggregated to the post level using max-pooling to preserve salient signals.
- Post-level feature vectors are averaged to create mean vectors for the concept and non-concept sets.
- Cosine distance ( $d_{cos}$ ) is calculated between these means. A larger distance indicates that the model encodes the concept in a direction distinct from unrelated text, signifying higher representational clarity.

Key Results

1. Latent Feature Visualization (UMAP)

The UMAP projections reveal distinct differences between models trained on the original UMD data versus the augmented data:

UMD Model: Features associated with under-represented topics (e.g., immigration, perfectionism) appear as scattered, sparse points, suggesting limited topic coverage and strong superposition (features responding to multiple phenomena).
Augmented (AUG) Model: The same topics exhibit a larger number of highlighted feature directions that form compact, coherent clusters. This indicates a shift toward monosemanticity, where specific feature directions respond more selectively to target psychological concepts.
Interpretation: Topic-aware augmentation increases the availability and consistency of topic-selective directions in the latent space, reducing entanglement for under-represented factors.

2. Quantitative Separability (Cosine Distance)

Table 1 in the paper summarizes the cosine distance between topic means and the non-concept baseline.

Significant Improvements: Augmentation leads to substantial increases in directional separation for several key topics. The largest gains are observed for Relationship Problems (+0.334), Financial Crisis (+0.310), and Family Issues (+0.266).
Broad Improvements: Consistent improvements (approx. +0.18 to +0.22) are seen for Anxiety, Racism, Anger, Bullying, and Immigration.
Exceptions:
- Hopelessness: Shows a slight decrease (-0.017), likely due to strong semantic overlap with other risk factors (e.g., depression), limiting its ability to be separated as an independent direction.
- Unemployment: Shows a decrease (-0.079), but this is attributed to a "ceiling effect," as the topic already exhibited very high separation in both models (0.82–0.90).

Key Contributions

The paper makes three primary contributions to the field of NLP interpretability in mental health:

Mechanistic Analysis: It moves beyond performance metrics to analyze the internal activation space of suicide detection models using overcomplete sparse autoencoders to extract interpretable feature directions.
Feature Specialization Analysis: Through UMAP visualizations and top-activating sample analysis, it demonstrates that topic-aware augmentation leads to more structured representations where specific features align with distinct psychosocial topics, forming coherent clusters.
Geometric Separability Framework: It introduces a quantitative framework using cosine distance to measure the clarity of psychological concept representation, demonstrating how augmentation reshapes the internal feature structure to better isolate risk factors.

Significance and Claims

The authors claim that their findings suggest topic-aware augmentation does more than improve classification performance; it fundamentally alters the internal geometry of the model to be more interpretable and robust.

Structured Representations: Augmentation helps the model form clearer, more distinct internal representations for under-represented psychosocial risk factors (e.g., immigration, family issues), reducing the reliance on spurious correlations with over-represented concepts.
Safety and Transparency: By promoting monosemanticity for specific risk factors, the models become more transparent. This allows for better auditing of what the model is "looking for" and supports safer, more accountable deployment in high-stakes mental health contexts.
Nuanced Polysemanticity: The paper notes that while monosemanticity is desirable for clarity, some polysemanticity (shared features) is natural and beneficial for overlapping concepts. The goal of augmentation is not to eliminate all overlap but to reduce harmful interference where unrelated content triggers the same risk signals.

The study concludes that understanding the geometric structure of learned representations is essential for diagnosing gaps in topic coverage and ensuring that suicide detection models generalize effectively across diverse psychological contexts.