Assessing Heterogeneity in Sentiment Changes in Text-Based Counseling: Latent Class Trajectory Analysis

Online text-based counseling services are becoming increasingly popular. However, their text-based nature and anonymity pose challenges in tracking and understanding shifts in help-seekers' emotional experience within a session. These characteristics make it difficult for service providers to tailor interventions to individual needs, potentially diminishing service effectiveness and user satisfaction.

This study aimed to identify distinct within-session sentiment trajectories among help-seekers in online text-based counseling and examine key variables associated with trajectory membership.

A total of 6207 counseling sessions were randomly extracted from an online text-based counseling service in Hong Kong. A latent class trajectory analysis of help-seekers' in-session sentiment was conducted using a growth mixture model (GMM) to identify latent groups of help-seekers exhibiting specific sentiment trajectories. Sentiment scores of help-seeker messages, labeled by ChatGPT, served as the primary variable for trajectory modeling. Subsequently, a multinomial logistic regression was performed to identify variables associated with class membership.

The GMM identified 3 distinct sentiment trajectories as the best fit: (1) steady improvement (1171/6207, 18.9%), (2) deterioration (1119/6207, 18.0%), and (3) dip-then-rebound (3917/6207, 63.1%). Compared with the Dip-Then-Rebound Class, help-seekers in the Deterioration Class were more likely to report suicidal ideation (OR=1.28, 95% CIs 1.07-1.52, P=.006), present with family (OR=1.56, 95% CIs 1.19-2.08, P=.002) or physical health-related concerns (OR=1.67, 95% CIs 1.02-2.74, P=.04), have an unknown gender status (OR=1.32, 95% CIs 1.04-1.67, P=.02), access the service through the anonymous channel (OR=1.30, 95% CIs 1.03-1.63, P=.03), depart from the session prematurely (OR=9.76, 95% CIs 8.33-11.36, P<.001), and have shorter session durations (OR=0.77, 95% CIs 0.71-0.84, P<.001).

We identified 3 distinct trajectories of help-seekers' in-session sentiment. Identifying the most likely trajectory at an early stage in the session could potentially help counselors adjust their approaches, thereby improving the effectiveness of text-based counseling and enhancing help-seeker satisfaction.