Open Collections

Laricheva, Maria

University of British Columbia

In the past two decades, researchers in counselling and psychotherapy have been increasingly applying natural language processing (NLP) methods to identify signs of mental illnesses in patients, assist counsellors with coding, and analyze the underlying processes in therapy. However, one of the challenges for applying NLP in counselling is how to prevent or minimize biases, such as racial or gender bias, introduced from training data or pretrained models. The biased results based on NLP research might lead to negative consequences for patients and counsellors, such as an inappropriate treatment plan. Unfortunately, this challenge has not been addressed in counselling research. In this thesis, I reviewed the common practice of NLP applications in counselling and psychotherapy and investigated the issue of gender bias in the counselling contexts. This thesis consists of two studies, which are presented in Chapters 2 and 3, separately. Chapter 2 is a scoping review of NLP applications for counselling and psychotherapy that utilize conversational data. This study summarized trends, strengths, and challenges in publications over the past three decades and examined the limitations and gaps of these publications, including bias issues. Chapter 3 investigated gender bias in an NLP model trained on historical counselling data. It compared two neural network models and proposed a new benchmark corpus for bias detection. The contribution of this thesis is twofold. First, our scoping review of NLP applications in counselling is the first study to summarize the methodological and bias issues. Given the consequential implications of bias, bias issues were not addressed appropriately in the counselling research, with a lack of discussion on the possibility of bias and the generalizability of the model’s results. This finding motivated our second study, in which we investigated how gender bias could be introduced by historical training data and explored how mask language modelling could serve as a ground for bias evaluation. We hope that our studies will draw researchers’ attention to the issues of bias in NLP models and their applications in counselling.

Text

2023-04-18

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/84232

Measurement, Evaluation and Research Methodology

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/