Article Preview
TopIntroduction
Rumors, unverified or speculative information circulated among the public, can significantly impact beliefs and behaviors, especially on sensitive topics like health and politics (Chierichetti et al., 2011). With the rise of online media, rumors spread easily, undermining trust and creating societal disruptions. Thus, it is crucial to verify information online. Figure 1 shows the example of some comments below on a website (Ma et al., 2016).
Figure 1. Examples of Rumor Information
Therefore, in recent years, the academic community's increasing interest in addressing the urgent need for effectively combating misinformation and rumors in information dissemination has led to continuous attention on pre-trained language models (PLMs) in the field of rumor detection. Previous research has primarily focused on feature extraction based on PLMs, as illustrated in Figure 2a. Figure 2a shows the entire process of model fine-tuning in rumor detection tasks. Task-specific fine-tuning involves adding headers specific to the task (such as rumor detection headers) to PLMs and fine-tuning the corresponding header layers to train and optimize the entire model. While these methods have yielded some results, the fine-tuning of PLMs for specific tasks, like rumor detection, has become prohibitively expensive with the advancement of PLMs. To tackle this challenge, we implemented the Prompt Learning method and integrated it with the task of rumor detection. Figure 2 shows the details of the model’s Prompt Learning.
Figure 2. Fine-Tuning (a) and Prompt Learning (b)
Prompt Learning activates the knowledge within PLMs using natural language prompts (Liu et al., 2023). The input text is passed through the prompt function and output into a prompt template containing the prediction mask. These templates are spliced with the input text and passed to PLM after predicting the classification labels. The Verbalizer module maps this prediction to the label set and can discern the authenticity of the information. Prompt Learning, striving to establish a closer connection between the pre-training task and the target task, similar to the direct execution in PLMs, more effectively utilizes the knowledge within PLMs, thereby enhancing the performance of rumor detection tasks. It is noteworthy that Prompt Learning excels in resource-limited environments, providing robust support for rumor detection tasks.
Additionally, to overcome the challenge of low data resources in rumor detection, we propose an approach for rumor detection, namely prompt-based learning for rumor detection (PLRD). This method embeds the rumor detection task into a prompt-based learning framework by leveraging prompt templates. Simultaneously, the introduction of T5 (Raffel et al., 2020) model-generated prompt templates aids in performing rumor detection tasks, enhancing the performance of PLM. The T5 model is a text-to-text transformer-based model developed by Google, capable of performing various natural language processing tasks, including text summarization, translation, question answering, and more.
We validate the effectiveness of the PLRD method through extensive experiments on the X and Weibo datasets (Ma et al., 2016). In a few-shot setting, our approach is both efficient and allows for a more competitive language model in a full-scale environment. Unlike PLM fine-tuning approaches, learnable prompts perform rumor detection directly and comprehensively through pre-training tasks to better mine text corpus information. Specifically, the model introduces two innovations: