Increasing accuracy an‎d reducing costs in stress detection from text using Active Learning

7/9/2025 12:00:00 AM

In recent years, the widespread use of social media an‎d the massive generation of textual data have opened new opportunities for analyzing individualsʹ mental states using artificial intelligence. One critical task in this domain is the timely detection of psychological stress, which can play a significant role in preventing more severe disorders such as depression. Given the importance of early stress detection, this study aims to develop an intelligent model for detecting stress in Persian texts using an active learning approach. However, the lack of efficient models for stress detection in Persian remains a major challenge. This gap highlights the need for appropriate dataset collection, labeling strategies, an‎d learning models tailored to the characteristics of the Persian language. Such limitations have hindered the development of effective text-based psychological analysis systems, underscoring the necessity for a localized an‎d practical solution. The proposed solution consists of two distinct approaches for data labeling an‎d model training. In the first approach, textual data collected from Twitter is fully annotated by human experts an‎d then used to train a deep neural network. In the second approach, an active learning framework is employed: a large language model is used for initial labeling, an‎d a confidence score is applied to identify challenging samples for manual annotation. This confidence score is also leveraged throughout the active learning process to focus on difficult examples, thereby enhancing the quality of the training set an‎d significantly reducing the number of non-informative samples. Ultimately, the comparison between the two approaches revealed that active learning not only significantly improves model accuracy but also greatly reduces the financial, time, an‎d human costs of data annotation. Nonetheless, two main challenges remain in this framework: first, selec‎ting the optimal confidence threshold, which depends on the data characteristics an‎d may vary under different conditions; an‎d second, the reliance on large language models for initial labeling, which can occasionally lead to incorrect labels even at high confidence scores.

مدل تشخيص استرس از متن فارسي , پردازش زيان طبيعي , يادگيري فعال , يادگيري عميق , هوش مصنوعي

Persian text-based stress detection model , natural language processing , active learning , deep learning , artificial intelligence

Hamid Nabilou

MohammadReza Jahed Motlagh