ارائه رويكردي براي تشخيص عواطف در سطح جنبه به صورت تك برچسبي در متن كوتاه فارسي

Social networks an‎d online platforms have provided a space for millions of users worldwide to interact an‎d share their views an‎d emotions on various social issues. Analyzing emotions expressed in these virtual spaces is of great importance for various groups such as policymakers, manufacturers, governments, an‎d users. In recent years, attention to Aspect-Level Emotion Analysis has increased as a more precise approach compared to overall sentiment analysis. This method identifies emotions associated with different aspects of a topic rather than assigning a single emotion to the entire text. Despite advances in this field, aspect-level emotion detection for the Persian language, especially in short texts, faces numerous challenges. The lack of stan‎dard datasets, structural complexities of the Persian language, an‎d the need to determine the dominant aspect for practical applications are among these challenges. Additionally, most existing systems are only capable of detecting three simple sentiment classes (positive, negative, neutral) an‎d do not cover the broader spectrum of human emotions. In this research, an innovative architecture called AspectEmo is presented for single-label aspect-level emotion detection in Persian short texts. The proposed system is based on the combination of two models, T5 an‎d EmoDetector, which operates in three stages: automatic aspect extraction, emotion detection for each aspect based on Plutchikʹs eight-emotion model (joy, trust, fear, surprise, sadness, disgust, anger, anticipation), an‎d dominant aspect determination using an algorithm that considers emotional intensity, aspect position in the text, an‎d category importance. To address the dataset scarcity challenge, a dataset containing 600 samples for Persian an‎d English was generated an‎d reviewed using Claude AI. eva‎luation results show that the proposed system achieved an F1-Score of 76.73% for English an‎d an F1-Score of 71.24% for Persian. Compared to the only similar existing work in English, the proposed method showed a 17.29% improvement in aspect extraction an‎d a 29.83% improvement in the combined task of aspect extraction an‎d emotion detection. This research represents the first comprehensive effort for aspect-level emotion detection in Persian an‎d provides promising results for future developments.

6

دكتر فرشته خوشنام

163580