Files

Abstract

While it is widely accepted that clients’ level of self-disclosure has important implications for treatment outcomes and therapeutic relationship, it is difficult to measure. Extant assessment tools fail to capture the fluctuating nature of self-disclosing behaviors, their sensitivity to context and vulnerability to subjective biases. This study sought to address this issue by utilizing a quantitative self-disclosure method assessing distinct linguistic features. Forty-eight participants aged 18-35 were recruited to complete two dyadic, 45-minute conversations with a stranger co-participant. The conversation topics they received either induced high or low self-disclosure. Logistic regression models in combination of natural language processing techniques including a word-count approach (LIWC) and a word-embedding approach (BART) were constructed based on participants’ conversation content to classify high vs. low-level of self-disclosure. A logistic regression model built upon LIWC categories achieved 89.66% accuracy, and the one using pre-trained BART model retained 48.28% prediction accuracy. Incorporating demographic variables did not affect the model built on LIWC but improved the accuracy to 62.07% for the BART approach. Results provides preliminary evidence in support of algorithms that assess language content to predict high versus low-levels of self-disclosure. Limitations were discussed that future studies might improve model performance by recruiting a larger, more heterogenous participant sample to discuss a wider range of topics.

Details

Actions

PDF

from
to
Export
Download Full History