يك مدل ميان‌دامنه‌اي براي تشخيص موضع چندهدفه‌ي متن‌هاي كوتاه

In the last decade, the expansion of social media and the growth of their use has turned these media into a valuable source of information. Automatic extraction of information from views and opinions expressed in social media is one of the important tasks in the field of natural language processing (NLP). Stance detection is one of these tasks, the purpose of which is to identify the favor, against or neutral position of the author of a message towards a target. A target is something or someone that a stance refers to. For texts with more than one target, the subtask of multi-target stance detection is proposed. In this subtask, the model must be able to detect the stance of all targets. By using a single-target model for all targets, previous researches have used these models to detect multi-target stances. One of the neglected cases in previous researches is the domain of targets. In this research, in order to improve efficiency and reduce the dependence of the model on the domain of training data, a domain is assigned for each target and the proposed model organizes received text, target, and domain in the form of one input. The additional knowledge that enters the model through the domain causes the model to pay more attention to the context discussed in the text and the target and makes the relationship between them more visible. In addition, the presence of domain in the inputs eliminates the need to train separate models for different domains. The eva‎luation of the proposed model shows a significant improvement in efficiency compared to previous researches. The minimum growth rate of 0.63% of the F1 measure indicates the success of the current research in increasing efficiency. Especially, the minimum growth of 10.11% of the F1 measure obtained on the unseen data, shows the good generalizability of the proposed model.

6