Adapting Mental Health Prediction Tasks for Cross-lingual Learning via Meta-Training and In-context Learning with Large Language Model

TL;DR

The paper tackles cross-lingual mental health prediction in a low-resource language (Swahili) by proposing two complementary approaches: a model-agnostic meta-learning (MAML) framework to enable rapid cross-language adaptation, and an in-context learning pipeline using large language models (LLMs) with carefully designed prompts. Empirical results show that meta-trained models outperform standard fine-tuning baselines in macro-F1 scores and demonstrate robust cross-lingual transfer, while LLM-based prompting reveals that English prompts generally drive stronger performance, with Swahili prompts providing notable gains when language-specific prompting is well crafted. The work contributes a practical, data-efficient pathway for deploying mental health prediction in underrepresented languages and highlights the potential of combining meta-learning with LLM prompting for real-world cross-lingual NLP tasks. Limitations include reliance on translated Swahili data and evaluation on a single low-resource language, suggesting future work on multiple languages and integrated meta-learning/LLM frameworks.

Abstract

Timely identification is essential for the efficient handling of mental health illnesses such as depression. However, the current research fails to adequately address the prediction of mental health conditions from social media data in low-resource African languages like Swahili. This study introduces two distinct approaches utilising model-agnostic meta-learning and leveraging large language models (LLMs) to address this gap. Experiments are conducted on three datasets translated to low-resource language and applied to four mental health tasks, which include stress, depression, depression severity and suicidal ideation prediction. we first apply a meta-learning model with self-supervision, which results in improved model initialisation for rapid adaptation and cross-lingual transfer. The results show that our meta-trained model performs significantly better than standard fine-tuning methods, outperforming the baseline fine-tuning in macro F1 score with 18\% and 0.8\% over XLM-R and mBERT. In parallel, we use LLMs' in-context learning capabilities to assess their performance accuracy across the Swahili mental health prediction tasks by analysing different cross-lingual prompting approaches. Our analysis showed that Swahili prompts performed better than cross-lingual prompts but less than English prompts. Our findings show that in-context learning can be achieved through cross-lingual transfer through carefully crafted prompt templates with examples and instructions.

Figures (5)

• Figure 1: An overview of the proposed framework: we use English as the source and Swahili as the target language. The meta-train stage transfers from the source to the target languages, while the meta-adaptation further adapts the model to the target language.
• Figure 2: LLM in-context settings. We use psychological data and social media datasets to construct input-output pairs. We create training and test splits, comparing zero-shot and few-shot settings models' prediction performance.
• Figure 3: An example of an instruction based input-output pairs for the depression severity prediction task using four prompting strategies. The task is formatted as a natural language sequence. Each input contains an instruction, instance and task specific prompts.
• Figure 4: Fine-tuning and meta-training on mental health tasks. We used training sets to train on English, Arabic and Swahili then fine-tuned aggregated data, varying training instances. Blue lines indicate mBERT initialised and purple XLM-R initialised.
• Figure 5: Performance accuracy results for the models in three learning scenarios using the three prompt settings: Swahili, Cross-lingual and English, across four mental health tasks.