Detection of Conspiracy Theories Beyond Keyword Bias in German-Language Telegram Using Large Language Models

Overview

• The paper addresses the challenge of detecting conspiracy theories in German Telegram messages during the COVID-19 pandemic.
• It compares the performance of supervised fine-tuning approaches using BERT-like models with prompt-based approaches using large language models like LLaMA2, GPT-3.5, and GPT-4.
• The researchers used a dataset of around 4,000 messages without relying on keyword filters, which can introduce bias into the models.

Plain English Explanation

The paper looks at ways to automatically detect conspiracy theories in German messages on the Telegram messaging app during the COVID-19 pandemic. [https://aimodels.fyi/papers/arxiv/rumour-evaluation-very-large-language-models] Typically, this kind of detection relies on supervised machine learning, where the model is trained on a set of labeled examples. However, creating that training data can be difficult and time-consuming, as the content is often harmful.

The researchers tried two different approaches: [https://aimodels.fyi/papers/arxiv/surprising-efficacy-fine-tuned-transformers-fact-checking]

1. Supervised fine-tuning: They took a BERT-like language model and fine-tuned it on their dataset of German conspiracy theory messages. This achieved an F1 score of around 0.8, which is comparable to recent models trained on English data.
2. Prompt-based approaches: They also tested using large language models like LLaMA2, GPT-3.5, and GPT-4 in a zero-shot setting, where the model is given a custom definition of conspiracy theories and asked to classify the messages. The best-performing model was GPT-4, which achieved an F1 score of around 0.8 without any additional training.

The researchers found that both approaches can be effective at detecting conspiracy theories in this context. [https://aimodels.fyi/papers/arxiv/adapting-fake-news-detection-to-era-large] They also showed that their supervised model could adapt to changes in the data over time, maintaining an F1 score of around 0.7.

Technical Explanation

The paper compares two approaches for detecting conspiracy theories in German Telegram messages:

1. Supervised fine-tuning: The researchers fine-tuned a BERT-like model (specifically, a German version of DistilBERT) on a dataset of around 4,000 Telegram messages collected during the COVID-19 pandemic. They did not use any keyword filters, which can introduce bias. The fine-tuned model achieved an F1 score of around 0.8 for the positive (conspiracy theory) class, comparable to recent models trained on English datasets.
2. Prompt-based approaches: The researchers also tested using large language models like LLaMA2, GPT-3.5, and GPT-4 in a zero-shot setting. These models were given a custom definition of conspiracy theories and asked to classify the messages without any additional training. The best-performing model was GPT-4, which achieved an F1 score of around 0.8 for the positive class.

[https://aimodels.fyi/papers/arxiv/chatgpt-vs-media-bias-comparative-study-gpt] The researchers also evaluated their supervised model's ability to adapt to temporal shifts in the data, simulating a real-world scenario where the model needs to handle changes in the distribution of conspiracy theories over time. They report an F1 score of around 0.7 in this setting, demonstrating the model's robustness.

Critical Analysis

The paper presents a valuable contribution to the challenge of detecting conspiracy theories in non-English social media content. [https://aimodels.fyi/papers/arxiv/stance-detection-social-media-fine-tuned-large] However, there are a few potential limitations and areas for further research:

• The dataset, while larger than many previous studies, is still relatively small, with only around 4,000 messages. Expanding the dataset size and diversity could help improve the models' performance and generalization.
• The paper does not provide much insight into the specific types of conspiracy theories present in the dataset or how the models perform on different subtypes. Analyzing the model's strengths and weaknesses across various conspiracy theory categories could yield useful insights.
• The prompt-based approaches rely on the language models' ability to understand and apply the provided conspiracy theory definition. Further research could explore ways to optimize the prompts or incorporate additional contextual information to improve the models' performance.

Overall, the paper presents a solid methodological approach and promising results, but there is still room for further exploration and refinement of the techniques for detecting conspiracy theories in diverse, multilingual social media settings.

Conclusion

This paper addresses the important challenge of automatically detecting conspiracy theories in German Telegram messages during the COVID-19 pandemic. The researchers compared the performance of supervised fine-tuning approaches and prompt-based approaches using large language models, demonstrating the effectiveness of both methods.

The supervised fine-tuning approach achieved an F1 score of around 0.8 for the positive (conspiracy theory) class, comparable to recent models trained on English datasets. The prompt-based approach, using GPT-4 with a custom conspiracy theory definition, also achieved an impressive F1 score of around 0.8 in a zero-shot setting, without any additional training.

These findings suggest that both supervised and prompt-based techniques can be leveraged effectively for detecting conspiracy theories in non-English social media content. The researchers also showed that their supervised model could adapt to temporal shifts in the data, maintaining robust performance over time.

Overall, this work contributes valuable insights and techniques for addressing the challenge of conspiracy theory detection, particularly in multilingual and evolving social media contexts.