طبقه‌بندي متون فارسي در آزمون حافظه سرگذشتي به دو طبقه خاص و بيش‌كلي‌گرا

Psychiatric disorders are significant medical issues that affect an individualʹs functioning and may lead to disruptions in daily activities and a reduced quality of life. One method for identifying these disorders is the eva‎luation of individuals’ memory. Autobiographical memory consists of a person’s life experiences, and assessing this memory can be effective in diagnosing disorders such as depression and anxiety. The Autobiographical Memory Test (AMT) is used to eva‎luate this type of memory. In this test, individuals are presented with a set of emotionally positive, negative, or neutral words and are asked to recall memories associated with these words. The recalled memories are then classified into four categories: specific memories, extended memories, categorized memories, and semantically associated memories. To enhance and accelerate the classification process, research has been conducted on automating memory classification using language models and machine learning methods in languages such as English, Japanese, and others. The present study aims to localize the Autobiographical Memory Test for the Persian language. Given the complexities of Persian, including its extensive vocabulary and diverse verb structures, classifying memories in this language is more challenging. The goal of this research is to develop a solution that can automatically classify the recalled memories from the AMT into a binary classification: specific memories (Class 1) and non-specific memories (Class 0). To achieve this objective, a higher-level approach to memories must be adopted to extract common features between specific and non-specific memories. In other words, the aim is to reduce dependency on the vocabulary of the memories in the classification process and ensure that classification occurs independently of word variations. The use of Persian Grammer and sentence embedding design can help achieve this goal. Moreover, processing embedded sentence vectors is simpler than processing word-level embeddings due to their lower dimensionality and constraints. The research findings indicate that the proposed classifier, designed with embedded vectors, outperforms the TookaBERT-based language model fine-tuned with textual memory data. The proposed classifier can automatically predict the memory class without reliance on specific words and has achieved superior results compared to other models. The classifier, eva‎luated using different metrics, achieved an accuracy of 86%, precision of 87%, recall of 87%, and specificity of 83%. Other classifiers were also trained using a translated dataset of English autobiographical memories for comparison and eva‎luation. The combination of different models, such as the proposed classifier, a neural network, and the 5NN model, was tested on the dataset, slightly improving the results but not significantly impacting overall performance. The results of the proposed classifier demonstrate a considerable advancement in automating the classification process of autobiographical memory test responses in Persian compared to pre-trained language models.

6

120186