Exploring the validity of sentiment analysis in psychotherapy

What happens if you apply the Multilingual Language Model Toolkit for Sentiment Analysis (XLM-T), “a transformer-based NLP model derived from the Cross-lingual Language Model based on RoBERTa”, to psychotherapy transcripts? Eberhardt et al. (2024) investigate. Here’s a slightly simplified Table 1, showing correlations between positive and negative sentiment and a patient-reported emotions scale. Green gives between-patient correlations and pink within-patient across sessions. Note the wide confidence intervals.

Eberhardt, S. T., Schaffrath, J., Moggia, D., Schwartz, B., Jaehde, M., Rubel, J. A., Baur, T., AndrΓ©, E., & Lutz, W. (2024). Decoding emotions: Exploring the validity of sentiment analysis in psychotherapy. Psychotherapy Research.

Harvey at PwC

LLM-driven text analysis is becoming a norm, allowing people to process huge volumes of text they wouldn’t otherwise have the capacity to do. Although outputs can be checked, the large volume of inputs processed means there are fundamental limits on how comprehensively analyses can be checked.

PwC announced yesterday that it is trialling the use of Harvey, built on Chat GPT, to “help generate insights and recommendations based on large volumes of data, delivering richer information that will enable PwC professionals to identify solutions faster.”

They say that “All outputs will be overseen and reviewed by PwC professionals.” But what about how the data was processed in the first place…?