حدیث بشیری

ارائه ­دهنده:

حدیث بشیری

  استاد راهنما:

دکتر حسن نادری

  هیات داوران:

دکتر بهروز مینایی
 دکتر عین‌اله خنجری
دکتر آزاده شاکری
 دکتر صادق علی‌اکبری


 زمان۲۴ اردبیهشت ماه ۱۴۰۴

  ساعت: :۱۲:۳۰
 

مکان: اتاق سمینار کارشناسی ارشد طبقه سوم


 

 چکیده پایان نامه :
 

Abstract

Nowadays, many individuals use social media to express their opinions and views. Given the growing importance of these platforms, they can serve as a valuable source for collecting public feedback. Sentiment analysis, which focuses on understanding people’s attitudes toward various entities and topics, assigns a score to each opinion (positive, negative, or neutral) and provides a deeper insight into public sentiment. This study aims to evaluate the popularity of events by analyzing user feedback. Assessing the sentiment polarity toward specific events can play a significant role in social, political, and economic decision-making. In addition to offering an overall understanding of event popularity, this information can also be used to assess public reactions and identify communities with shared viewpoints. These findings may also serve as a basis for friend recommendation systems. The proposed framework in this study consists of two main components: event detection and sentiment analysis. In the first stage, events are identified using an event detection method, and in the second stage, sentiment analysis is performed on the data associated with each event. Finally, a popularity score is assigned to each detected event. The novelty of this research lies in both components. For sentiment analysis, a method called SyntaPulse is introduced. By combining a lexicon-based approach with an unsupervised machine learning model, it addresses key challenges in sentiment analysis, including textual ambiguity, domain dependence, sarcasm detection, and the lack of labeled data across various domains. This framework achieves high accuracy and strong performance across diverse datasets. In the event detection component, the framework introduces innovations such as dynamic bandwidth adjustment based on local data density, the use of multivariate distance, adaptive kernel density estimation, and an improved Louvain-MOMR method for community detection, which enhance the effectiveness of event identification. For final evaluation, due to the lack of suitable labeled data, three datasets—FACup, SuperTuesday, and USElection—were sentiment-labeled using two pre-trained models. The results show that the proposed framework performs well in both components and can accurately estimate the popularity of events. Furthermore, separate evaluations of the sentiment analysis and event detection components indicate that, in most cases, they outperform existing state-of-the-art methods.