Using large language models for extracting and pre-annotating texts on mental health from noisy data in a low-resource language

Introduction

Artificial intelligence, particularly large language models (LLMs), is increasingly being explored as a tool for various mental health care tasks within both psychiatry and psychology, as is shown in multiple recent reviews (Guo et al., 2024; Haque & Rubya, 2022; He et al., 2023; Hua et al., 2024; Li, 2024; Obradovich et al., 2024; Volkmer, Meyer-Lindenberg & Schwarz, 2024; Yang et al., 2023). These tasks can be broadly categorized into several areas: automatic detection of psychiatric disorders, psychological conditions, or personality traits from diverse data sources; generation of treatment recommendations for mental health practitioners; and support for mental health care seekers through automatic dialogue systems or conversational agents (CAs) designed to provide advice, counseling, and psychotherapeutic sessions.

The success of AI systems in addressing these tasks, however, heavily depends on the availability and quality of training data, which poses a significant challenge, particularly for low-resource languages and societies. While diagnostic systems can, in principle, utilize any data types as long as the correct diagnoses are present in the training sets, dialogue systems require question-answer text pairs whose relevance and quality have been validated by experts. In cases where conversations between mental health seekers and providers are unavailable, one approach is to label texts from open sources, such as online user-user or user-professional forums. However, the proportion of high-quality or even relevant texts in these data sets is so low that manual assessment would be highly inefficient, if not impossible. This situation creates a demand for automatic pre-filtering to eliminate irrelevant texts and for pre-annotation or pre-classification of the remaining texts into meaningful categories, such as discussions of specific disorders and psychological conditions. Once enriched in this manner, text collections can be more efficiently hand-labeled in terms of their topics, the quality of responses, and other important features. Social media data provide an opportunity to train models for pre-annotation tasks since they often contain user labels that, albeit loosely, are able to indicate certain text classes, such as topic, sentiment or communication purpose.

In this work, we address the task of pre-filtering and pre-annotating of noisy online data for the subsequent training of mental health care conversational agents by exploring the capabilities of LLMs operating in zero-shot classification mode. One potential application of such a system is to support users who have been diagnosed with a specific psychiatric disorder or who suspect they have one and are seeking to develop a coping strategy. For efficient labeling of data to train such a model, it is desirable to automatically pre-classify texts into disorder-specific categories. To achieve this, we utilize data from pre-selected forums and online communities that include source posts describing a user’s situation, attributed by the author to a specific psychiatric diagnosis, along with multiple responses of varying quality and relevance. Text categories (psychiatric disorders) are derived from thread names or source post hashtags and are used as ground truth for evaluating LLM quality and for model fine-tuning.

Zero-shot classification (ZSC) is a machine learning approach that enables a model to recognize new classes without any prior examples. This capability arises because large language models (LLMs) are pre-trained on vast amounts of data, making them relatively accurate in predicting or imitating human responses to queries formulated in natural language. For instance, LLMs can provide reasoned and detailed answers to questions about a text’s topic, specific issues, reasoning methods, and tone, resembling human-generated responses. However, the quality of such classification is often constrained by the lack of relevant texts in the data used to pre-train a given LLM. To address this issue, LLMs are further trained (fine-tuned) on smaller, more relevant datasets, thereby leveraging both extensive general-knowledge data and more focused domain-specific data.

In our work, we explore several approaches to classifying user texts into categories that reflect specific psychiatric disorders. First, we perform pure zero-shot classification (ZSC) into these categories. Second, we precede this ZSC with an additional round of ZSC aimed at filtering out irrelevant texts—those that do not belong to any of the target categories. Third, we fine-tune the LLMs using our dataset. All experiments are conducted on the data retrieved from the Russian language social media.

Related Work

In this section, we primarily focus on experimental studies and reviews that address the challenges of data availability, quality, and generation in the domain of machine learning for psychiatric purposes. We predominantly reference works from 2023 to 2024, a period marked by significant methodological breakthroughs in LLMs and a corresponding increase in research employing these models for mental health tasks. In this domain, datasets can be categorized into two major types:

• Data about individuals. This includes information on actual or potential mental healthcare recipients and healthy individuals, encompassing their medical records and various digital traces, such as social media posts or data on social networking and app usage. When annotated, these data are labeled with either medically confirmed or self-ascribed diagnoses and are used to train models for classification tasks, such as diagnosis, screening, and risk assessment.

• Data containing mental healthcare seeker queries and responses. The latter can be either human-generated, AI-generated, or both, and are labeled according to criteria such as correctness, professionalism, usefulness, or empathy. These datasets are used to develop dialogue systems.

Most experimental studies, as well as review articles, focus on the efficiency of various algorithms for solving psychiatric tasks, once the data is already available and its quality is considered acceptable. However, several recent review articles highlight data shortages and other data-related issues as significant challenges to the development of LLM applications in psychiatry (Obradovich et al., 2024; Guo et al., 2024; Volkmer, Meyer-Lindenberg & Schwarz, 2024). They include unavailability due to formal privacy regulations protecting relevant data (Obradovich et al., 2024), reluctance of large commercial data owners to share the data (Obradovich et al., 2024), privacy risks for individuals whose data are already in use (Chung, Dyer & Brocki, 2023), lack of non-English datasets and multilingual datasets, especially with expert annotation (Guo et al., 2024), lack of information on or understanding of the datasets when they are available (Guo et al., 2024), and insufficient data quality leading to errors in diagnosis or risk estimation. Other researchers also point at obstacles on the way of creation of high quality datasets. Volkmer, Meyer-Lindenberg & Schwarz (2024) mention uncertainties about creating adversarial examples and difficulties with labeling multi-modal data, which is common in psychiatry. Demszky et al. (2023), although discussing LLM use in psychology, draw attention to a highly relevant problem of low agreement between experts when evaluating the psychological usefulness of AI-generated responses.

According to the available reviews (Hua et al., 2024; Guo et al., 2024), most datasets used in psychiatry-relevant machine learning (ML) are constructed for classification tasks, with text-generation tasks being the second most common. Manual construction is more prevalent than supervised and automatic construction, and datasets annotated by trained laypersons outnumber those annotated by experts. Nearly all datasets are in English, with a few in Chinese and several other languages, ranging from several hundred to a few dozen thousand samples. Social media platforms, especially Reddit and Twitter (now X), are the most common sources of these datasets.

While social media data offer the advantage of availability, they also present certain drawbacks. For instance, they contain little data on older age groups, which might hinder classification tasks (Guo et al., 2024). Similarly, if social media data are used to train psychiatric chatbots, the latter may inherit social prejudices prevalent among internet users whose texts were used for training (Chung, Dyer & Brocki, 2023). Additionally, Aich et al. (2024) observe that the use of large social media datasets often leads to reliance on non-clinical judgments as ground truth in diagnostic tasks, whereas datasets with clinically verified diagnosis labels tend to be very small.

As current proliferation of health-related chatbots is driven primarily by commercial logic, their architectures and the training data often not disclosed. Consequently, existing comparative studies of conversational agents typically do not provide information that would allow attribution of a specific CA’s success to the data it was trained on. Some of these studies are qualitative. For instance, Haque & Rubya (2022) investigated ten commercial mental health chatbots based on over 6,000 consumer reviews analyzed qualitatively. They concluded that the bots generally provided emotionally satisfying communication and immediate help, although they could foster excessive attachment and were not perceived as sufficient replacements for human psychotherapy. Similarly, Martinengo, Lum & Car (2022) found that the nine conversational agents they studied were capable of supportive communication and guiding users in mood-boosting activities. This conclusion was based on the analysis of agent-user interactions by trained assessors using structured criteria.

Other studies employ quantitative approaches. He et al. (2023) conducted a meta-analysis of the efficiency of conversational agent interventions (CAI) based on 32 studies. They demonstrated that CAI improved the symptoms of most investigated psychiatric disorders, including depression, anxiety, and stress, in the short term, but not in the long run. None of these works addressed which CAs were better and why. To the best of our knowledge, Li et al. (2023) is one of the few studies that provides such insights. Based on a meta-analysis of 15 studies involving over 1,500 participants, they found that stronger effects on mental health outcomes were achieved by generative CAs compared to retrieval-based ones, by multi-modal CAs compared to text-only ones, and when delivered via mobile devices compared to desktops.

The number of works proposing solutions for the automatic or semi-automatic annotation of mental health data is very limited. As a nascent domain, it is currently developing primarily within non-medical natural language processing (NLP) and for simpler tasks such as named entity recognition (NER), sentiment analysis, and topic and spam detection. Some studies have tested the quality of existing large language models (LLMs) against human annotation, reporting comparable quality for the simplest NLP tasks (Li, 2024; Nasution & Onan, 2024), but lower quality for more complex tasks, such as recognizing six basic emotions (Nasution & Onan, 2024).

The general approach of more advanced methods involves designing human-LLM collaborative systems. These systems typically start by training an annotator LLM on a small amount of human-annotated data. Subsequently, small subsets of LLM-annotated samples are re-annotated by humans and re-submitted as training data for a new round of model training. The final quality of such systems is tested on small held-out datasets that were not involved in any stage of training. This approach was followed by Kholodna et al. (2024) in the NER task; like Nasution & Onan, they experimented with low-resource languages. Wang et al. (2024) took this a step further by introducing an evaluator LLM into the loop, which identified potentially poorly annotated samples to be sent for human re-annotation instead of random samples. Huang et al. (2024) went even further by proposing that LLMs be prompted to develop programs for annotation rather than generating annotations themselves. The goal was to make further annotation cost-free; however, the reported quality varied greatly depending on the task.

The work in the mental health domain most closely related to the task of automated annotation is by Yang et al. (2023). The authors prompted LLMs to determine whether a user was likely to suffer from a given psychiatric disorder based on the user’s text and then to explain the response. By experimenting with zero-shot, few-shot, and emotion-enhanced prompts, they achieved an F1 score between 0.6 and 0.84 on binary classification datasets, but only 0.44 on T-SID multi-class dataset. Further, they compared human and LLM-generated evaluations of the explanations for the obtained labels and found that the overall correlation between human and LLM evaluations was around 0.5. Aich et al. (2024) worked with transcripts of pre-diagnostic interviews conducted by clinicians with patients suffering from bipolar disorder, schizophrenia, and healthy individuals. Interview texts were professionally annotated based on five speech markers indicative of these disorders, such as clarity, focus, and social appropriateness. An LLM was trained to interview individuals and annotate their speech according to these five markers. Fu et al. (2023), while developing a system to help non-professionals conduct online psychological interventions, proposed an idea that could be useful for mental health data annotation. After asking non-professionals to respond to user queries, they prompted an LLM to evaluate the quality of these lay responses. These machine evaluations were further assessed by experienced counselors, who graded 78% to 97% of the LLM-generated evaluations as good, depending on the criterion. This idea of an evaluator LLM is similar to that employed by Wang et al. (2024).

Finally, it is worth noting the absence of relevant works, datasets, or solutions for the Russian language. Although the general-purpose Russian-language LLM, GigaChat (Krestnikov, 2024), has been rapidly developing, no Russian-language mental health CAs exist, except for a Telegram bot proposed by a Kazakhstani team (Omarov et al., 2023). Other CAs developed in countries with Russian-speaking populations are, in fact, English-based. There is one dataset constructed for the early detection of suicidal signals from social media texts (Buyanov & Sochenkov, 2022) and a study predicting well-being and depression risk from multimodal social media and mobile device usage data (Panicheva et al., 2022). Thus, at present, there is no foundation for developing human-LLM annotation systems in the Russian language.

The research gap, thus, can be defined as the lack of cheap fully automated pre-filtering and pre-annotation approaches that could be loosely verified with the existing non-clinical labels and further supplied to professional psychiatrists for expert annotation. This annotation could subsequently be used in mass human-LLM labeling systems leading to mental health CA development. Currently, the described lack is particularly severe for Russian language.

Zero-shot classification

In this work, we have chosen to apply zero-shot classification (ZSC) because this machine learning technique is well-suited for low-resource tasks. As noted above, in this approach, a model can classify data into multiple classes it has never encountered before, i.e., without any specific training examples. This is made possible through the use of auxiliary information; for instance, models can be trained to understand the relationships between different classes and can further transfer this understanding to identify other (unknown) classes based on their similarity (Zhang, Xiang & Gong, 2016). ZSC has been successfully applied in various fields, including natural language processing (NLP) and computer vision (Fu et al., 2018). For example, in NLP, models have been trained to understand semantic relationships between words, which subsequently enabled the models to perform sentiment analysis (Blitzer, Dredze & Pereira, 2007), classify texts by topic (Alcoforado et al., 2022), and execute other tasks without specific training examples. An overview of zero-shot classification models and the tasks where ZSC is used is provided in the articles by Wang et al. (2019), Chen et al. (2021) and Pourpanah et al. (2023).

Based on the works of Fu et al. (2018) and Yin, Hay & Roth (2019), the following modes of ZSC operation can be distinguished.

• 1)

  Statement-label mode: This is a straightforward classification mode in which an LLM is tasked to classify statements (or texts) in accordance with pre-determined labels. The essence is that the user can specify the names of the categories by which the data should be classified.

• 2)

  Statement-hypothesis mode: This mode can be viewed as a ZSC-adapted application of natural language inference (NLI). NLI is an approach used to test a broad range of hypotheses about text pieces (termed premises), extending beyond simple classification to include assumptions about the logical connections between two phrases, among other things. The NLI task is formulated as binary, where the outcome is either a confirmed or rejected hypothesis. In the NLI adaptation for the classification task, the model tests the hypothesis about whether a text belongs to a given class and determines its probability before yielding a binary response (yes/no). This mode is particularly well-suited for cleaning text collections of irrelevant data.

• 3)

  Fine-tuning mode: In this mode, the LLM can be further trained if a relevant dataset is available, enabling it to classify texts more effectively. This can be done in two ways: standard fine-tuning that involves adding a final layer to the LLM and retraining the entire model, and NLI training.

Each of the large language models tested in this work has been run in all three modes.

Materials and methods

Computer experiments

The following experiments were conducted on the datasets described above.

Results

While the experimental results in terms of overall accuracy for all models are summarized in Table 4, Appendices A–D contain more detailed results on performance as measured with multiple metrics (precision, recall, F1 for each class and generalized over all classes as average and as Macro, for each model and dataset).

Conclusion

In this work, we analyzed the performance of four zero-shot large language models on noisy text data from the mental health domain, with a primary focus on discussions of coping strategies for psychiatric disorders. The goal was to evaluate the ability of LLMs in different modes to pre-filter and pre-classify noisy online data, thereby reducing the workload for annotators in low-resource languages. This task was formulated specifically for the mental health domain and for data that could potentially be used to train disorder-specific mental health conversational agents in the future.

In the following discussion of our results, we compare them to the work of Yang et al. (2023), which is the only known study where LLMs were tested in zero-shot mode to classify psychiatry-related texts. It is important to note that a direct comparison is not possible since the datasets used by Yang et al. (2023) are English-language, noise-free, and mostly two-class. The most relevant three-class dataset, T-SID, contains tweets categorized by high, low, and no suicide risk.

Not unexpectedly and consistent with the findings of Yang et al. (2023), we observe that non-fine-tuned LLMs in ZSC mode perform poorly. The accuracy achieved on our data is 10% lower on the T-SID dataset and does not exceed 0.22. Additionally, we find no improvement in classification quality when using lemmatized data compared to non-lemmatized data, indicating that this preprocessing step can be safely omitted. The superior performance of multilingual models compared to English-only models can be attributed to their ability to directly process Russian-language texts, as they include Russian in their training data. In contrast, translating Russian texts to English may result in information loss. Therefore, translation should only be recommended for languages that lack language-specific LLMs.

Further, we demonstrate that filtering enhances the ability of LLMs to pre-classify data, albeit only by a few percentage points, which falls short of our initial expectations. This limited improvement may be attributed to the fact that the filtering model itself was not fine-tuned, whereas it was fine-tuning that contributed to the quality increase most significantly at the subsequent stage (again, in line with Yang et al. (2023)). The maximum accuracy we achieve with fine-tuned models is 0.64, which is nearly 20% higher than the fine-tuning approach proposed by Yang et al. (2023), although it remains lower than some domain-adapted models explored by Yang et al. (2023). A promising avenue for future research here would be to combine a fine-tuned filtering model with subsequent fuzzy classification, both using domain adaptation. Fuzzy approach could be particularly beneficial, as mental healthcare seekers’ discussions of their coping strategies often encompass multiple existing or alternative diagnoses.

Contrary to our expectations, a dramatic increase in accuracy, approximately 40% to 45%, achieved through model fine-tuning, does not significantly differ between standard fine-tuning and the more advanced Natural Language Inference (NLI) approach. However, the latter results in a substantial increase in computation time, roughly sixfold. Future improvements in NLI fine-tuning could be achieved through more sophisticated hypothesis engineering and incorporating human input, as recently demonstrated for non-psychiatric NLP tasks by Kholodna et al. (2024), Wang et al. (2024), and Huang et al. (2024). It is important to note that straightforward usage of human annotation for reinforcement learning might be not advisable with social media data, and more nuanced approaches to providing context for annotators should be used. Additionally, experimenting with intermediary models to identify potentially poorly annotated samples for further human re-annotation could be beneficial. Existing studies in the mental health domain offer mixed results: such models have shown poor quality in Yang et al. (2023) but high quality in Fu et al. (2023), indicating the need for further experimentation.

In summary, our research makes several significant contributions. We have conducted the first exploration of the potential for LLMs to pre-annotate mental health data in a language for which no relevant annotated datasets currently exist, and for which extensive annotating resources are not anticipated in the near future. Through our investigation of various approaches, we have identified some unnecessary or undesirable steps, such as lemmatization and the use of translated datasets. More importantly, we have pinpointed promising steps, the most notable of which is domain-specific model fine-tuning. Finally, our dataset and fine-tuned models together present a unique resource for the low-resource task of developing a Russian-language conversational agent capable of maintaining disorder-specific mental health counseling dialogues. While our models can be employed as components of such a CA, the dataset can be used for the creation of a vector database for model training and prompt optimization.

Supplemental Information