Inferring Depression and Its Semantic Underpinnings from Simple Lexical Choices

Line Kruse, Roberta Rocca, Mikkel Wallentin

TL;DR

This paper introduces a novel, simple language task called the Demonstrative Choice Task (DCT) that can infer depression symptom severity with surprising accuracy. By analyzing how individuals choose between "this" and "that" for isolated words, the study reveals semantic patterns linked to depression, offering a less intrusive way to assess mental states and understand individual differences in depressive experiences.

What Problem Does This Paper Address?

Diagnosing and understanding depression is complex. While standard symptom scales are crucial, they often fail to capture the nuanced, underlying experiential states that contribute to the disorder. Depression is characterized by a heightened and often negative focus on the self, but the precise semantic and cognitive mechanisms driving this self-focus and its manifestation in language remain an area of active research. Existing methods for inferring depression from language can be intrusive or require large amounts of data. The authors sought a simpler, more indirect method to identify individuals with depressive symptoms and to characterize the semantic underpinnings of these states, aiming to bridge the gap between reported symptoms and internal experiences.

Methods

The core of this research is the Demonstrative Choice Task (DCT). In this task, participants are presented with individual English nouns, one at a time, without any contextual information. For each noun, they must choose between two response options: a proximal demonstrative ("this") or a distal demonstrative ("that"). The key hypothesis is that this choice reflects not just physical proximity, but also emotional or experiential proximity to the referent word.

The researchers developed logistic regression models to classify participants into "control" or "depression" groups based on their DCT responses. They compared several models:

  • mDCT: A model using only DCT responses.
  • mDCT+Demo: A model that includes DCT responses along with demographic features (gender and age).
  • mGenderAge: A baseline model using only demographic features.
  • mRandomBaseline: A control model trained on randomly shuffled outcome labels to assess chance performance.

Model performance was evaluated using out-of-sample classification accuracy and ROC AUC scores. Additionally, post hoc semantic analyses were conducted to understand which words and which semantic features were most predictive of depression. This involved correlating the model's learned weights with semantic features from the NRC-VAD lexicon (e.g., valence, arousal, dominance, joy, sadness, fear).

Datasets

The study utilized two independent, cross-sectional datasets collected via the online platform Prolific.

  • Study 1: 775 participants (after exclusions).
  • Study 2: 879 participants (after exclusions).

Participants were native English speakers aged 18 and over. Data was collected through the DCT and the 9-item Patient Health Questionnaire (PHQ-9) for self-reported depression symptom severity. No new benchmarks were introduced; the study leverages existing semantic lexicons.

Key Insights and Ideas

  • "This" vs. "That" as a Window into the Self: The study demonstrates that simple choices between "this" and "that" for isolated words can reveal deep-seated semantic and emotional associations, particularly concerning self-focus, which is a hallmark of depression.
  • Emotional Proximity Reflected in Lexical Choice: The DCT goes beyond physical space, suggesting that demonstrative choices can map onto an individual's internal, emotional, or experiential space.
  • Depression Skews Semantic Proximity: Individuals with higher depression scores tend to associate "negative" words (e.g., sadness, fear, disgust) with proximal demonstratives ("this"), suggesting a closer, more internalized experience of these negative concepts. Conversely, positive words might be associated with distal demonstratives.
  • A Less Intrusive Assessment Tool: The DCT offers a potential avenue for assessing depression symptoms indirectly, bypassing direct self-report questions that might be subject to bias or social desirability.
  • Semantic Profiles of Depression: The research moves beyond simply classifying individuals to characterizing the semantic characteristics of their experiential states, opening doors to understanding individual differences within the spectrum of depression.

Key Results

  • Classification Accuracy: The mDCT and mDCT+Demo models achieved classification accuracies ranging from 60% to 66% across the two studies, significantly outperforming chance and demographic-only models.
  • Predictive Power of Negative Affect: Demonstrative choices for words scoring high on negative valence (sadness, fear, disgust, anger) were consistently strong predictors. Specifically, proximal choices for negative words predicted depression, while distal choices predicted the control group.
  • Replication Across Samples: The findings were robust and replicated across two independent samples, strengthening their validity.
  • Semantic Feature Correlations: Post hoc analyses revealed that semantic features like sadness, fear, disgust, and anger were negatively correlated with classification weights (meaning proximal choices for these words predicted depression), while joy, trust, and valence showed positive correlations (meaning proximal choices for these words predicted lower depression scores in Study 1).

Limitations and Future Work

  • Not a Diagnostic Tool: The authors emphasize that the DCT models, while accurate, are not a substitute for clinical diagnosis. They cannot definitively confirm the presence of clinical depression and are limited by the accuracy of the PHQ-9 self-report measure they are validated against.
  • Sample Characteristics: The studies used online samples, and future work should aim to validate these models in samples of clinically diagnosed patients.
  • Task Context: While the task is designed to be context-free, transient mental states or task context (e.g., the day of the week) could influence responses. Future research could explore how to leverage these transient dynamics.
  • Expanding Semantic Space: The current analysis relies on a specific set of semantic features. Expanding the semantic feature space and exploring other word categories (e.g., social, body, money) could provide a more nuanced understanding of individual differences.
  • Optimizing Predictive Performance: While the current accuracy is promising, the authors suggest that adapting the model, perhaps by assigning higher weights to specific word categories or replacing non-informative items, could further optimize predictive performance.
  • Longitudinal Studies: Investigating these patterns in longitudinal DCT studies could reveal within-subject dynamics relevant to depression.

Summary updated

Abstract

Spatial demonstratives are highly frequent linguistic universals, with at least two contrastive expressions (proximal ("this") vs. distal ("that")) indicating physical, social, or functional proximity of the speaker to the referent object. Recent evidence based on the Demonstrative Choice Task (DCT), in which participants couple words with a spatial demonstrative with no context provided, suggests that demonstrative use is also indicative of experienced or emotional proximity to the self in an imagined mental space. As depression is characterized by increased and maladaptive focus on the self, the DCT may be a simple and reliable way to elicit behaviors that enable inference on the presence of severe depressive states and allow descriptions of the semantic characteristics of individual differences in such states. In two independent cross-sectional studies, including 775 and 879 participants, respectively, we showed that DCT-based classification models reliably capture semantic characteristics of experiential states that are predictive of self-reported depression symptom severity, as measured by PHQ-9. In both samples, DCT classifiers outperformed baseline models and replicated semantic patterns of negative affect previously observed to be associated with depression. This indicates that the paradigm captures semantic characteristics of the experiential states underlying depression symptoms and may be used to map individuals along a broad semantic space, potentially providing novel insights into individual differences in depressive states.
BibTeX
@inproceedings{kruse2024inferring,
  author = {Line Kruse and Roberta Rocca and Mikkel Wallentin},
  title = {{Inferring Depression and Its Semantic Underpinnings from Simple Lexical Choices}},
  booktitle = {Depression},
  year = {2024},
  doi = {10.1155/2024/3010831},
}
Figure