Abstract
Traditional dictionary-based text analysis often fails to capture the nuanced, population-specific linguistic markers of mental health. This study demonstrates that combining Linguistic Inquiry and Word Count (LIWC) with unsupervised Latent Dirichlet Allocation (LDA) significantly improves the prediction of neuroticism and depression in college students. By identifying latent themes such as somatic complaints and vegetative symptoms, the hybrid model outperforms standard lexicons and matches or exceeds the performance of clinical psychologists in screening tasks.
Introduction
Clinical psychology relies heavily on patient language for diagnosis. However, traditional assessments like the MMPI are time-consuming and depend on accurate self-reporting, which can be hindered by stigma or lack of self-awareness. While automated tools like LIWC have been used to link word usage to psychological states, they rely on pre-defined, static categories. This research explores the use of data-driven methods to identify subtle, context-specific linguistic markers of mental distress that fixed dictionaries might overlook.
Methodology
The study compares a standard baseline against an augmented model using two primary datasets: a Neuroticism Dataset (6,459 essays) and a Depression Dataset (124 essays).
Feature Extraction
- LIWC (Baseline): A lexicon-based approach counting words in 64 categories (e.g., negative emotion, pronouns).
- LDA (Topic Modeling): Latent Dirichlet Allocation was used to discover latent themes within student essays. This unsupervised model identifies clusters of words that frequently co-occur.
- Combined Model: The researchers tested the union of LIWC and LDA features (with topic counts of 20, 30, 40, and 50) using linear regression to predict psychological scores.
Key Insights: Capturing the "Missing Middle"
A significant advantage of LDA is its ability to identify population-specific themes. While LIWC includes a general "body" category, LDA discovered more granular themes highly relevant to clinical depression, such as "Somatic" and "Vegetative/Energy" levels.
These themes capture specific student experiences, such as "Homesickness," which are diluted in broader dictionary categories. The results suggest that top-down linguistic categories and bottom-up data-driven themes provide complementary signals for mental health assessment.
Results
Neuroticism Prediction
The addition of topic features significantly improved the Pearson correlation ($r$) over LIWC alone. The hybrid model achieved its peak performance with 40 topics.
As shown in the table, LIWC+LDA (40 topics) reached an average $r = 0.459$, a statistically significant improvement over the $0.413$ achieved by LIWC alone ($p < .02$).
Depression Prediction and Human Comparison
The improvement was even more pronounced for depression, where adding topic features boosted the correlation from $r = .288$ (LIWC only) to $r = .416$. In a 30-item test set, the LIWC+LDA model achieved an F1 score of 0.50, outperforming two out of three practicing clinical psychologists and matching the third. The automated model tended to be more conservative, favoring precision over recall.
Discussion and Limitations
While promising, the study is limited by a small sample size for the depression dataset, which required oversampling. Additionally, while the model showed high precision, a screening tool might benefit from higher recall to avoid missing potential cases. Future work will explore more sophisticated models like Supervised LDA and hierarchical topic regression to further refine prediction accuracy in real-world clinical settings.
Summary updated