Topic-Oriented Controlled Text Generation for Social Networks

Currently, advances in text modeling by Pre-trained Language Models (PLM) enable machines to generate texts that are fluent and human language-specific. However, in the social network scenario, the generated text still has the issues of not being able to recognize sarcastic opinions, not conforming to social network language conventions, not containing topic-related knowledge information, and degrading some attributes when controlled by multiple attributes, which cannot control the stance, style, and topic of the generated text accurately. Besides, there is a lot of false and malicious news on the Internet, jeopardizing the security of the network and big data. To address these challenges, we defined and analyzed the attributes of topic texts and then proposed a PCTG-X model for single attribute discriminator control of PLM decoding based on prompt learning in order to control the stance, style, and topic attributes, respectively. We collected topic datasets from social platforms such as Twitter, and used some publicly available social text datasets to design model evaluation metrics and comparison experiments. The experiments show that PCTG-X has a significant improvement in the control of three single attributes of stance, style and topic compared with the benchmark method.