چكيده به لاتين
Nowadays, social media is a platform for freely expressing and sharing opinions and thoughts. This
leads to the fact that by analyzing the data available on social media, a broad and comprehensive
perspective on various users’ opinions and sides about different topics could be gained. These topics include political, economic, social, and cultural issues. In Natural Language Processing, stance
detection is the process of automatically recognizing the side and stance of a given text about a specific target.
In natural language processing tasks, the way text data is preprocessed significantly affects the performance of the trained model. In this research, seven different levels of preprocessing are introduced and examined. Additionally, to find the architecture of the stance detection model, the idea of
Neural Architecture Search (NAS) was inspired. In this method, the model architecture is divided
into four main parts, a search space is defined for each part, and adaptive search algorithms are used
to design the final architecture. The best proposed model ultimately utilizes BERTweet as the encoder and a CNN classifier. The proposed architecture achieved an F1-Score of 74.47%, showing
a 19.97% improvement over the Baseline model. Furthermore, the proposed method ranked third
among 19 participants in a climate change stance detection event. Additionally, due to the lack of
training data for different topics, stance detection without training data was also investigated. This
approach, which uses large language models and prompt engineering, introduces four approaches
based on different prompt types. Then, the performance of the proposed prompts was compared
with other methods for stance detection without training data. The introduced approach achieved an
F1-Score of 57.33%, showing a 2.03% improvement over similar approaches.
