Research On The Stance Detection Method Of Network Topics Oriented To Curriculum Ideology And Politics

The development and reform of ideological and political education has always been valued in our country.In 2020,the "Guidelines for the Ideological and Political Construction of College Curriculum" issued by the Ministry of Education clearly pointed out the concept and connotation of "Curriculum Ideological and Political",and proposed to use specific courses as the carrier in the teaching process to effectively integrate "ideological and political elements" into curriculum teaching new forms of ideological and political education in all aspects of education.Since then,major colleges and universities have begun to integrate "ideological and political elements" into the teaching of "curriculum ideological and political".However,integrating "Ideology and Politics" elements into actual teaching process requires specific events or topics,and there are numerous hot topics and discussion texts related to the connotation of "Ideology and Politics" generated daily from social media platforms.Using stance detection technology to extract hot topics related to "Ideology and Politics" from social media platforms is a feasible approach.Since stance detection is based on the topic,it is not realistic to collect all topics in the actual application process.Therefore,research on zero-shot standpoint detection is a good complement.Based on self-built network ideological and political topic datasets and public datasets,this paper conducts research on the two focus directions of stance detection based on specific topics and zero-shot stance detection.Existing topic-specific stance detection research does not take into account that comment texts with the same stance on the same topic often discuss some of the same or similar content,which itself contains certain stance information.To address this issue,this paper proposes a keyword-aware contrastive learning model for topic-specific stance detection.The model takes into account representative content within the same topic and extracts important words that contain stance tendencies,while preserving the original word order to form a keyword sequence.Furthermore,based on the keyword sequence,the paper proposes the Keyword Perception Contrastive Learning approach to encourage the similar vector representation of keywords expressing the same stance on the same topic in the model,thus enhancing the model’s ability to express stance information.Finally,the experimental results show that the stance classification performance of the model is improved by 1.44%,1.25% and 0.87% respectively in three public datasets(Sem2016,COVID-19 and WT-WT)compared with the baseline model,and increased by 1.71% on the self-built network ideological and political topic dataset.Further experiments show that keyword-aware contrastive learning can capture the similarity between keywords expressing the same stance under the same topic,thereby improving the classification performance of the model.Existing zero-shot stance detection models learn text-to-topic stance representations from context without simultaneously focusing on general representations of stances expressed in text(topic-invariant features).To address this issue,this paper proposes an adversarial learning model based on a dynamic scheduling strategy for zero-shot stance detection.The model is able to learn the stance representation in the context while dynamically combining the adversarial learning to extract the general representation,which improves the quality of the model learning the stance representation from the context.Furthermore,existing zero-shot stance detection evaluation methods cannot accurately evaluate the model’s performance when faced with the same test topic under different training conditions.Starting from a new evaluation perspective,this paper fuses six public datasets and constructs two simulated datasets.Finally,the experimental results show that compared with the compared baseline model,the model improves by 2.08% and 5.44% on the two simulated datasets,and improves by 1.31% on the self-built network ideological and political topic dataset,and the model achieves higher performance improvement in topics with lower experimental results.Compared with the previous work based on adversarial learning,the model proposed in this paper does not rely on a large number of unlabeled examples to assist training,and can be more easily extended to other datasets and has better applicability.