Predicting Depression in Screening Interviews from Latent Categorization of Interview Prompts

Alex Rinaldi, Jean E. Fox Tree, Snigdha ChaturvediACL 2020 · OpenReview

Alex Rinaldi, Jean Fox Tree, Snigdha Chaturvedi. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 2020.

Introduction

Automated depression detection from clinical interviews is a critical area of research, yet it often struggles with two primary challenges: context-dependency and model interpretability. In a standard screening interview, the significance of a patient's response is inextricably linked to the interviewer's prompt. For instance, a brief "Yeah" might indicate simple agreement or, conversely, a lack of engagement depending on whether it follows a direct question or an encouraging backchannel. Traditional deep learning models often treat dialogue as a flat sequence, overlooking these structural nuances. This paper introduces the Joint Latent Prompt Categorization (JLPC) model, which explicitly accounts for conversational context to improve both predictive performance and transparency.

Methodology: Joint Latent Prompt Categorization (JLPC)

The JLPC model is designed to mimic the hierarchical structure of a conversation by categorizing interviewer prompts into latent groups and aggregating patient responses accordingly.

The JLPC model architecture, showing the Category Inference Layer and Category-Aware Response Aggregations.

The architecture consists of several key components:

  1. Category Inference (CI) Layer: A feed-forward layer assigns each interviewer prompt ($P_{im}$) to one of $K$ latent categories ($h_{im}$). This allows the model to differentiate between types of prompts, such as "Starters" or "Backchannels," without explicit manual labeling.
  2. Category-Aware Aggregation: Patient responses ($R_{im}$) are grouped based on the inferred category of the preceding prompt. This mechanism enables the model to learn category-specific weights for linguistic features.
  3. Entropy Regularization: To ensure the model discovers distinct and meaningful categories, an entropy penalty is applied to the latent category distributions, preventing the model from collapsing all prompts into a single category.
  4. JLPCPost Variant: This version incorporates an average representation of all prompts into the final decision layer, capturing the overall "style" of the interview.

Data and Preprocessing

The model was evaluated using the Distress Analysis Interview Corpus (DAIC), which includes 170 transcripts of interviews between human participants and a virtual agent. A significant methodological contribution of this work is the identification and removal of "predictive" prompts. The authors observed that interviewers often ask specific follow-up questions (e.g., "How long ago were you diagnosed?") only when they already suspect depression. To prevent the model from "cheating" by identifying these giveaway questions, 31 types of predictive prompts were iteratively removed from the training data.

Results

The JLPC model demonstrates superior performance in identifying depressed individuals compared to standard baselines, including fine-tuned BERT models.

Comparison of F1 scores for depressed and non-depressed classes across various models, highlighting the superior performance of JLPCPost.

As shown in the results table, the JLPCPost variant achieved an F1 score of 0.440 for the depressed class, significantly outperforming the fine-tuned BERT baseline (0.362). This suggests that modeling the structural relationship between prompts and responses is more effective for this task than increasing raw model capacity alone.

Discussion and Insights

Beyond performance gains, the JLPC model provides psycholinguistic insights. The model's latent categories naturally aligned with human-interpretable conversational roles, such as "Starters" (e.g., "Tell me more") and "Backchannels" (e.g., "mhm"). Analysis revealed a "leveling off" effect: while depressed individuals initially show reduced social engagement, the gap narrows when interviewers use frequent backchannels, highlighting the therapeutic value of encouragement. Furthermore, the model identified that depressed individuals often use goal-oriented language (e.g., "wished," "mission") specifically in response to specific backchannels like "nice."

Despite these successes, the study is limited by the small size of the DAIC dataset (170 interviews) and its reliance solely on text. Future research could integrate vocal prosody and facial expressions, as well as explore temporal patterns in how patient language evolves over the course of an interview.

Summary updated

BibTeX
@inproceedings{rinaldi2020predicting,
  author = {Alex Rinaldi and Jean E. Fox Tree and Snigdha Chaturvedi},
  title = {{Predicting Depression in Screening Interviews from Latent Categorization of Interview Prompts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  url = {https://openreview.net/forum?id=7cBx1wAbFcb},
}
Figure