Towards Enhancing AI-Driven Mental Health Support with an Intelligent Counsellor Agent

Abstract

Recent strides in the development of conversational models for goal-oriented applications present substantial opportunities for augmenting human expertise and automating aspects of professional tasks. Among the most critical areas where such advancements could be transformative is the domain of mental health support—an arena of paramount importance to families, businesses, enterprises, and society at large. Mental health disorders present a substantial global burden, with over 50% of adults experiencing mental health issues at some point in their lives. Despite the widespread prevalence of these conditions, a staggering 70% of affected individuals worldwide do not receive treatment from professional healthcare providers. From 2011 to 2030, it is estimated that the global economic impact of mental disorders will amount to a loss of US $16.3 trillion in cumulative output, surpassing the economic losses attributed to cancer, chronic respiratory conditions, and diabetes. Mental health disorders among working-age adults are another pressing concern, with 15% of this population affected and a staggering 12 billion working days lost annually to depression and anxiety alone. The World Health Organization underscores the critical need for practical mental health support in occupational settings, highlighting its significant role in boosting individual’s confidence, productivity, and the capacity to maintain or secure employment. In this thesis, we explore how AI-driven solutions, such as chatbots and tools to enhance responses by mental health professionals, have the potential to provide accessible and efficient support for mental health issues. Unfortunately, prior work in this space is significantly limited due to the lack of quality data for training, with no existing large-scale conversational dataset present. To address this, we create the first large-scale mental health conversational dataset (BBMH) with over 36,000 threads of conversations sourced from a popular mental health forum. Further, to tackle productivity loss at the workplace, we introduce the task of building classifiers to filter workplace-related mental health posts. We create weak-labelers with pretrained language models and experiment with zero-shot prompting on state-of-the-art large language models to filter out a high-quality dialog dataset for mental health support at the workplace (BBWMH) with 5,240 threads and 49,238 utterances. We also build strong baseline dialog models using state-of-the-art large language models such as Gemma, Mistral, and Llama-2 for workplace-related mental health support. In addition to our work on mental health support, this thesis presents several methods by which artificial intelligence can be leveraged for social good, aligning with the objectives of the Project Angel initiative by IREL and IIIT Hyderabad. As part of our contributions to natural language understanding, we conduct an in-depth study of fine-grained hierarchical sexism classification. Our findings demonstrate the effectiveness of domain-specific pretraining, the application of focal loss, and multi-level training methodologies in addressing this task. Insights gained from this study have been instrumental in developing a workplace-related mental health post classifier, where we also address substantial class imbalance within the dataset. Mental health support is a highly complex task with a substantial body of psychological research accumulated over many decades, especially when help-seekers have mental health disorders. However, emotional distress, a common experience among the population regardless of the presence of diagnosed mental disorders, has been identified as a key area where intervention can be effective. Recent research highlights the benefits of Emotional Support Conversations (ESC) as a method to alleviate such distress and improve mental health outcomes. Previous methods for ESC condition on the conversational context to first predict a single support strategy and then generate an Emotional Support (ES) response. Unfortunately, these studies fail to imitate the intricate art of (a) dynamically understanding the evolving problem category distribution of the help-seeker and (b) responding using a combination of strategies known to be effective in addressing the predicted problem distribution. To address these limitations, we develop a psychologically grounded framework, Problem Identification and Strategy Matching for ESC (PRISM), which leverages Transformer-based encoder-decoder models to predict problem distribution (over ten types) and strategy distribution (over eight types) and generates effective Emotional Support responses. Unlike existing models that utilize coarse-grained strategy annotations and conversation-level emotion labels, training PRISM models requires fine-grained, probabilistic annotations for both support strategies and underlying problems. Hence, we contribute a novel dataset, ESConv++, curated using problem and strategy annotations from GPT-4. PRISM models are trained using ESConv++ to optimize four novel loss functions designed specifically for the ESC task. We conduct extensive ablation studies and human and automatic evaluations to evaluate our model and understand its various components. Our proposed model, PRISM, outperforms strong baselines on automated metrics by large margins and also performs best on human evaluation.