Natural Language Processing for Mental Health

In psychological therapies, understanding what works for whom and when remains one of the most enduring challenges in mental health research and practice. Today, however, we are better equipped than ever to address this question – thanks to transformative advances in both clinical science and AI. In this talk, I demonstrate how our interdisciplinary team of clinicians and AI researchers is bridging top-down, theory-driven approaches with bottom-up, data-driven AI methods to uncover the dynamics that drive therapeutic change.

I begin by outlining key challenges in conceptualizing, treating, and researching mental health. I then discuss significant theoretical shifts in clinical science – such as the transition from general treatment models to transtheoretical processes and from one-person to two-person psychology – that have paved the way for addressing these challenges. Building on these foundations, I illustrate how our team has leveraged theoretical innovations and advancements in multimodal analysis and AI to explore the dynamic processes within clients (intrapersonal dynamics) and between clients and therapists (interpersonal dynamics) that are linked to improved treatment outcomes.

Central to this effort is the development of temporally aware, multimodal AI methods designed to address key limitations in modeling temporality, complex reasoning, situational awareness, personalization, and the integration of verbal and non-verbal data. I conclude by discussing how this integrative approach can enhance diagnostic precision, support personalized interventions, and improve the overall effectiveness of mental health treatments.

When starting to formalize psychological constructs, researchers traditionally rely on two distinct approaches: the quantitative approach, which defines constructs as part of a testable theory based on prior research and domain knowledge, often using self-report questionnaires, or the qualitative approach, which gathers data mainly in the form of text and bases construct definitions on exploratory analyses. We present a new computational method that combines the comprehensiveness of qualitative research and the scalability of quantitative analyses to define psychological constructs from semi-structured text data. Using structured questions, participants are prompted to generate sentences that reflect instances of the construct of interest. We apply computational methods to calculate embeddings as numerical representations of the sentences, which we then run through a clustering algorithm to arrive at groupings of sentences as psychologically relevant classes. The method includes steps for measuring and correcting bias introduced by data generation and for assessing cluster validity through human judgment. We demonstrate the applicability of our method on an example from emotion regulation. Based on short descriptions of emotion regulation attempts collected through an open-ended situational judgment test, we use our method to derive classes of emotion regulation strategies. Our approach shows how machine learning and psychology can be combined to provide new perspectives on the conceptualization of psychological processes.

This two-part talk begins with a vision for HLAB focusing on (a) a more accurate representation of people, (b) LLMs for safe mental health therapy, and (c) eudaimonix content recommendation from AI. It continues with a series of experiments addressing the assumption that models trained on standard instruments are bounded by their predictive validity. We demonstrate cases where language-based assessments (LBAs) predict psychological outcomes better than the instruments they were trained on. Evidence includes: a historical review, a simulated experiment, an experiment with pseudo-observed data, and another with fully observed data predicting external criteria. These findings provide theoretical and empirical evidence that language-based assessments can more closely approximate true psychological states. Mechanisms by which these assessments outperform traditional tools are explored, highlighting the potential for AI-based language analysis to reshape psychological measurement.

The first part of the talk provides an overview of the NLP landscape for mental health, discussing the range of applications and the challenges faced by systems based on large language models (LLMs), with examples from the literature, especially regarding how these factors impact their suitability and applicability to mental health. There are many unresolved challenges, among others, regarding appropriate generation, temporal robustness, temporal and other forms of reasoning, and privacy concerns, especially when working with sensitive content such as mental health data. The programme of work I have been leading in the past five years consists of three core research directions: (1) data representation and generation, (2) methods for personalised longitudinal models and temporal understanding, (3) evaluation in real-world settings, with use cases in mental health. I will give an overview of the work in my group on these topics and conclude with a presentation of the evaluation platform for LLM-based systems that we have been developing within the AdSoLve project.

This presentation discusses a dynamic network model and research program supporting the delivery of personalized feedback to therapists at the start and throughout psychological therapy. The approach focuses on identifying the core individual elements of psychological distress and resources, while also enabling the quantification of multiple dimensions of distress (e.g., cognitive-behavioral, emotional, motivational, and interpersonal stressors) through the application of large language models. It addresses questions such as: How can psychological distress and resources be extracted, both qualitatively and quantitatively, from session transcripts, and how can they be integrated into clinical support tools and adaptive treatment planning for therapists? How can such models support the identification of patients at risk of treatment failure? And how can these models support clinical training? What are the technical and practical challenges that need to be addressed?

Traditional methods of psychiatric evaluation face ongoing challenges regarding reliability, objectivity, and scalability. The integration of language models offers a potential solution for assessing psychiatric symptoms and informing theory through digital biomarkers. In this talk, I will discuss findings on using linguistic markers to capture mental health symptoms from clinical conversations. In particular, I will introduce VISTA, a scalable method for capturing temporal flows, and apply it to emotions expressed by a sample of over 10,000 patients receiving digital mental-health treatment. I show how the resulting clusters relate to clinical outcomes. These findings highlight the opportunity to monitor patient outcomes using only linguistically captured emotions, especially when direct measurement of mental-health outcomes is infeasible.

I will discuss three provocations for the field of AI/NLP for mental health, guided by insights from empirical work in HCI and human-centered AI. My approach to this problem combines my disciplinary training in Media Studies and Computer Science. To unpack the transformative potential of human-centered AI, we’ll look at my group’s work in mental illness and online social systems (social media and generative AI) as examples. The goal of this is to inspire conversation among participants, encourage solutions, and spark interdisciplinary reflection.

Psychiatric interventions often focus on symptom alleviation, lacking the ability to fully capture the intricate dynamics of cognitive-affective states or tailor treatments to individual needs. From a dynamical systems perspective, though, the intervention can be observed in terms of an optimal control problem, where the ability to recover the dynamics based on continuous observation is key, as is the flexible implementation of control signals aimed at influencing cognitive-affective states – mirroring the severity and dynamics of mental disorders such as depression. Advances in natural language processing (NLP) and large language models (LLMs) offer transformative potential to address these limitations by enabling continuous, high-resolution monitoring of mental states through language, and further optimizing the processes of language-based therapies. This talk presents an integrative framework that combines physics-inspired dynamical systems theory with NLP. Language serves as a rich, temporally dynamic proxy for mental states, with LLMs facilitating the extraction of latent states and their temporal dynamics for predictive and interventional purposes. Optimized sampling rates, informed by ecological momentary assessment (EMA) and nonlinear modeling, enhance the detection of short- and long-term mood fluctuations. Furthermore, NLP-driven interventions, such as optimized psychotherapy and affective priming, can act as control inputs to reshape these trajectories, paralleling advances in neurostimulation that target pathological neural attractors. Early studies in our group show promising results in these directions, including those based on social media data (400,000 texts from more than 1,600 individuals over 6 years) and further experimental studies on optimizing interventions such as affective priming. This AI-informed neurocognitive dynamical systems framework paves the way for personalized interventions that stabilize resilient cognitive-affective trajectories, offering a paradigm shift from current approaches.

Recent years have seen a gold rush to replace people with AI agents in communication: they can serve as your therapist, your tutor, your financial advisor, and your interviewer. In this talk, I propose a contrasting vision: one in which AI supports humans in their communication while preserving their agency. Achieving this vision requires moving beyond the current transactional paradigm embodied by current generative AI systems, which are designed to fulfill the immediate goals of a single person, such as answering a question, solving a math problem, booking a flight, or (repeatedly) replying in character. To meaningfully support human–human communication without disrupting or supplanting it, an AI system must instead follow a proactive paradigm: it needs to decide when to intervene to offer support as the interaction unfolds, rather than wait to explicitly be prompted as AI agents and chatbots do today. In this talk, I present initial progress on AI technologies that enable such a proactive mode of operation and demonstrate them in the context of mental health crisis counseling. Data and code are available through ConvoKit.

Rating scales have shaped psychological research, but are resource-intensive and can burden participants. Large Language Models (LLMs) provide a tool for assessing latent constructs in text. This study introduces LLM rating scales that use LLM responses rather than human ratings. We demonstrate this approach using an LLM rating scale to measure patient engagement in therapy transcripts. Automatically transcribed videos of 1,131 sessions from 155 patients were analyzed using DISCOVER, a software framework for local multimodal human behavior analysis. Llama 3.1 8B rated 120 engagement items, averaging the top eight into a total score. Psychometric evaluation showed a normal distribution, strong reliability ($\omega =0.953$), and acceptable fit ($\mathrm{CFI}=0.968$, $\mathrm{SRMR}=0.022$), except $\mathrm{RMSEA}=0.108$. Validity was supported by significant correlations with engagement determinants (e.g., motivation, $r=.413$), processes (e.g., between-session efforts, $r=.390$), and outcomes (e.g., symptoms, $r=-.304$). Results remained robust across bootstrap resampling and cross-validation, accounting for the nested data structure. The LLM rating scale exhibited strong psychometric properties, demonstrating the potential of the approach as an assessment tool. Importantly, this automated approach uses interpretable items, ensuring a clear understanding of measured constructs, while supporting local implementation and protecting confidential data.

The global mental health crisis demands innovative approaches to scale high-quality care. While AI chatbots for therapeutic use are on the rise, I argue that research should develop scalable solutions for contexts where human support remains essential. I will present our work developing CARE, a Large Language Model (LLM)-based training system that empowers human counselors through practice with AI-simulated patients and feedback from AI mentors. In this talk, I highlight two research thrusts: technical challenges in training counselors with LLMs and evaluating the impact of LLM-based training with novice counselors. Advancing these research thrusts requires answering interdisciplinary questions at the intersection of natural language processing, mental health, and human AI interaction. How might we develop realistic LLM simulations of patients when privacy concerns restrict data access and domain-expert feedback is expensive? How can we mitigate the chance of generic or clinically-inappropriate LLM feedback? How can we conduct stage-appropriate user studies of LLM training systems that yield actionable insights to improve design? Our work advancing CARE contributes to responsible AI in mental health by developing new tools, algorithms, and empirical evidence for scaling counselor education via LLMs, thereby increasing the supply of well-trained human therapists to meet society’s growing demands.

Globally, there is a demand–capacity problem in mental health care. One in four individuals has a behavioral health disorder (substance use and/or mental illness), yet more than half of all behavioral health providers report no openings for new patients, and many report burnout with existing caseloads. Peer Recovery Workers (PRWs) have been cited as critical to improving access and engagement in behavioral healthcare and mitigating staffing shortages. However, high rates of burnout and turnover are well-documented challenges to implementing PRWs, often stemming from inadequate training and supervision. Artificial Intelligence (AI) could be an ideal solution for ensuring high-quality training and ongoing supervision of PRWs in a strained healthcare system with limited resources. In this talk, we review how peer competencies include many of the interpersonal skills foundational to effective psychotherapy (e.g., rapport-building, active listening, reflecting, validating, empathizing), and the value of scaling peer training in these competencies using AI. CARE is an AI-powered web-based platform to train and empower individuals in basic counseling skills. In CARE, individuals select and/or create AI patients to roleplay emotional support conversations. Using a Large Language Model trained by senior psychotherapy supervisors, CARE provides counselors with feedback on their responses to the AI patients using a multi-level structure that mirrors how supervisors provide feedback to trainees. Although lab experiments have established CARE’s proof-of-concept to improve the feedback quality provided to novice counselors, the program was not designed for PRWs whose scope of practice and training needs differ from those of traditional counselors (e.g., sharing lived experiences). Consequently, there may be a need to customize and pilot the use of CARE with PRWs. Mixed-method designs are suggested to collect systematic feedback from relevant stakeholders to guide customization of CARE for peers and then evaluate the feasibility and acceptability of the customized version with novice peers working in real-world clinical settings. This research program has the potential to address the demand–capacity problem by increasing the adoption of the PRW workforce in behavioral health settings and improving the quality of care they provide to patients.

In this talk, I present a Human–Computer Interaction perspective on the design of digital health interventions for mental health. Drawing on experience developing and evaluating a range of novel systems to support the delivery of mental health care, the talk considers the critical yet often ill-defined concepts of acceptability and engagement, before examining in more detail the HCI issues surrounding the use of machine learning technologies in this context, looking at a specific example relating to outcome prediction. The talk concludes with a brief consideration of more personalised interventions, including the use of LLM-based capabilities within a broader design space for Ecological Momentary Interventions, along with the associated design issues.

In this talk, the speaker discusses his NLP practices for mental health applications in China. Specifically, he developed tools with LLMs for mental state assessment, built chatbots to provide effective emotional support, and developed LLM models for simulating clients and therapists for the purpose of training and evaluation. Some of these tools have been deployed in real-world applications.

Mental health conditions exhibit extreme behavioral variability. A stunning example is Autism Spectrum Disorder (ASD), a catchall term covering a wide range of symptoms and varying degrees of communication difficulties. Only recently have we begun to understand the types and causes of ASD, and specialized interventions that work for every individual are still a far-off dream. AI, particularly large language models (LLMs), offers a unique opportunity to continuously learn from a user’s behavior, along with demographic information, patient history, and assessment results, which could serve as sandboxes for testing different (potentially novel) interventions before delivering the optimal one. In my talk, I discuss recent work that simulates user behavior in terms of what a user knows and does not know, and how to explain unfamiliar concepts in a lucid and personalized way from a cross-cultural communication perspective. These studies reveal that LLMs can learn to personalize, but they also tend to produce stereotypical, less variable responses than real users. LLMs often agree with responses from other LLMs, but not as much with human users, who align more strongly with each other. This insight suggests that culture can serve as an excellent prior for personalization, but systems must also continuously learn from unfolding behavior to achieve the best results. Such methods provide a promising direction toward individual mental health–relevant behaviors for intervention sandboxing and innovation.

Emotions provide critical cues into our health and well-being. They are particularly important in the context of mental health, where changes in emotion may signify changes in symptom severity. However, information about emotion and its temporal variation is often accessible only through survey methodologies (e.g., ecological momentary assessment, EMA), which can become burdensome over time. Affective computing technologies, such as automated speech emotion recognition systems, could provide an alternative, namely offering quantitative measures of emotion using acoustic data captured passively from a consented individual’s environment. However, these technologies are not without risk and can pose a potential for harm. There are critical ethical issues that must be thoughtfully considered. In this talk, I discuss our journey in affective computing for health modeling, presenting the design of these technologies alongside the ethical considerations that have shaped their development.

In this talk, I explore the critical privacy and security challenges of Large Language Models (LLMs). Users are increasingly turning to LLMs for sensitive and personal interactions, including advice, companionship, and counseling. This trend creates a “Personalization–Privacy Paradox,” in which the utility of personalized AI is in direct conflict with the need to protect user data.

The talk outlines concrete privacy threats, including membership inference, data leakage, and prompt extraction attacks. To combat these risks, a dual approach is proposed: Proactive Privacy (privacy by design) and Reactive Privacy (unlearning). I discuss recent results showing that differential privacy can reliably work for steering vectors, and that current unlearning evaluations are suboptimal – providing only a false sense of privacy. Finally, I present ongoing work on data poisoning and attribution methods for language models.

Speech is about more than content – it’s about what you say, when, and how. The latter two facets of spoken language – the “when” and the “how” – fall under the purview of computational paralinguistics. The use of para emphasizes contrast to classic linguistics. Paralinguistics concerns phenomena often neglected by traditional computational linguistics. In this talk, I provide a brief overview of the field, starting with motivational examples and a taxonomy of the relevant phenomena. I outline methodological approaches used in computational paralinguistics, from traditional feature-based analysis to modern foundation model–based methods that integrate speech and language processing. I also present a recent case study on using speech in the mental health context.

This working group, led by Iryna Gurevych (Technical University of Darmstadt, DE), focuses on how AI systems for mental health should be evaluated before they can be trusted in sensitive clinical and community settings. Building on analysis of recent NLP work on mental health, the group identified recurring gaps in current practice, including over-reliance on automated metrics, limited involvement of clinicians or people with lived experience, and insufficient attention to safety, equity, and context.

In response, the group is developing a taxonomy that distinguishes assessment, intervention, and support systems and links each aspect to appropriate evaluation dimensions. The resulting position paper will provide a shared, clinically grounded framework that allows AI researchers, clinicians, implementation scientists, and other stakeholders to speak a common evaluative language and move toward more reliable, interpretable, and socially responsible AI for mental health.

A second working group, led by Diyi Yang (Stanford University, US), examines the emerging use of AI-simulated patients as training tools for psychotherapy and counseling. The group starts from the long tradition of actor-based roleplay in clinical training and asks what it would take to extend these practices into scalable, AI-driven simulation environments. Key questions include how to balance conversational realism with pedagogical value and safety; how to represent long-term therapeutic trajectories versus short, skills-focused exercises; and how to ensure cultural, linguistic, and demographic diversity, including rare and edge-case presentations.

The group also considers data governance, risks of “overfitting” trainees or models to simulators, and methods for benchmarking educational utility across therapeutic orientations. The planned perspective paper will articulate design principles, evaluation standards, and a research roadmap for AI-simulated patients that are not only technically sophisticated but also ethically sound and transferable to real-world clinical practice.

The third working group, led by Dana Atzil-Slonim (Bar-Ilan University, IL), focuses on intrapersonal and interpersonal dynamics as central mechanisms of change in mental health. The group brings together clinical and AI researchers to ask which aspects of within-person experience and between-person interaction – such as emotional trajectories, synchrony, co-regulation, or conversational redirection – are most predictive of positive (or negative) outcomes, and how these phenomena can be modeled using modern AI methods.

The planned position paper will first synthesize clinical and methodological work on verbal and non-verbal dynamics at multiple temporal scales (from utterances to sessions and longer-term courses of care), and then review AI techniques for capturing temporal and multimodal patterns. Building on this, the group will identify promising intersections between the two literatures, highlighting opportunities to leverage advances in NLP and multimodal models to analyze longitudinal mental health datasets to identify intrapersonal and interpersonal dynamics predictive of positive outcomes.