TL;DR
This paper tackles the challenge of detecting conspiracy theories in German Telegram messages, moving beyond traditional keyword-based methods that introduce bias. The authors compare supervised fine-tuning of BERT-based models with prompt-based approaches using large language models (LLMs) like GPT-4 and Llama 2, finding that LLMs, particularly GPT-4, offer competitive performance with less reliance on labeled data.
What Problem Does This Paper Address?
The spread of conspiracy theories (CTs) online, especially during crises like the COVID-19 pandemic, poses a significant threat to informed decision-making and public trust. While automated detection methods exist, they often rely on supervised learning with human-annotated data, which is time-consuming and emotionally taxing. Crucially, many existing datasets and models are English-centric and heavily dependent on keyword filtering. This keyword bias limits their effectiveness by focusing only on explicit narratives and particular CTs, making them less applicable to real-world scenarios where CTs can be fragmented or subtly referenced. This research aims to address this gap by developing and evaluating methods for detecting CTs in German-language Telegram messages without relying on keyword filters.
Methods
The study compares two main approaches:
- Supervised Fine-Tuning: This involves taking pre-trained BERT-based models (specifically, German-language models like
deepset/gbert-base) and fine-tuning them on a custom dataset of German Telegram messages. The authors also explored further pre-training (retraining) on a larger in-domain corpus to improve performance. - Prompt-Based Classification (Zero-Shot and Few-Shot): This leverages the capabilities of large language models (LLMs) like GPT-3.5, GPT-4, and Llama 2. Instead of extensive fine-tuning, these models are given specific instructions (prompts) to classify messages.
- Zero-Shot: The LLM is asked to classify messages based on a definition of CTs provided in the prompt, without any prior examples.
- Few-Shot: The LLM is provided with a small set of example messages and their corresponding labels to guide its classification.
The researchers also investigated the impact of different prompt configurations, including the inclusion of a custom CT definition versus generic text or no definition at all.
Datasets
The primary dataset used is TelCovACT, a collection of approximately 3,663 German-language messages from public Telegram channels known for disseminating conspiracy theories related to the COVID-19 pandemic. This dataset was manually annotated for the presence of CTs, narrative components (actor, strategy, goal), and stance. Importantly, this dataset was collected without the use of keyword filters, addressing a key limitation of prior work.
The authors also created two "transfer datasets" by sampling messages from a wider range of Telegram channels and a later time frame to evaluate the temporal robustness of their models.
Key Insights and Ideas
- LLMs as a Viable Alternative: Large language models, especially GPT-4, can achieve performance comparable to fine-tuned BERT models, even in zero-shot settings, significantly reducing the need for extensive labeled data.
- Beyond Keywords is Crucial: The study's success with the TelCovACT dataset, collected without keyword filters, highlights the importance of moving beyond biased keyword-based approaches for more robust and generalizable CT detection.
- GPT-4's Superiority: GPT-4 consistently outperformed other LLMs and even rivaled the best fine-tuned model, particularly when provided with a custom definition of conspiracy theories.
- Fragmentation Matters: The detection performance of both fine-tuned and prompt-based models decreases when conspiracy narratives are highly fragmented (i.e., missing key components like actor, strategy, or goal), suggesting that narrative structure plays a role in detection difficulty.
- Zero-Shot Outperforms Few-Shot: Surprisingly, zero-shot LLM approaches, especially with GPT-4, performed better than few-shot approaches in this study, contradicting some prior expectations.
Key Results
- Supervised Fine-Tuning: The best fine-tuned BERT model (dubbed TelConGBERT) achieved an F1 score of ~0.79 for the positive class and a macro-averaged F1 score of ~0.85 on the test set. This performance is comparable to models trained on English, keyword-focused corpora. TelConGBERT also showed moderate transferability to new temporal data, with F1 scores around ~0.72 and ~0.67 on transfer datasets.
- Prompt-Based Classification:
- GPT-4: Achieved the best zero-shot performance with an F1 score of ~0.79 for the positive class and a macro-averaged F1 score of ~0.84, especially when using a custom CT definition. This performance was comparable to the fine-tuned TelConGBERT.
- GPT-3.5 and Llama 2: Showed competitive but generally lower performance than GPT-4. Llama 2, in particular, underperformed, likely due to its primary training on English data.
- Zero-Shot vs. Few-Shot: Zero-shot settings generally yielded better results than few-shot settings across the LLMs.
- Model Agreement: While TelConGBERT and GPT-4 performed comparably, their predictions disagreed on about 15% of the test data.
Limitations and Future Work
- LLM Costs and Proprietary Nature: The study acknowledges the high computational and monetary costs associated with using LLMs like GPT-4, as well as their proprietary nature, which can limit accessibility for smaller organizations.
- Few-Shot Performance: The consistently suboptimal performance of few-shot learning compared to zero-shot learning warrants further investigation into strategies for selecting in-context examples.
- Llama 2's German Performance: Llama 2's underperformance is attributed to its English-centric training. Future work could explore fine-tuning German-specific Llama 2 models or investigating other open-source German LLMs.
- Model Disagreement and Explanations: The authors note the interesting disagreement between models and suggest exploring qualitative analysis of LLM-generated explanations to understand these differences better.
- Dataset Expansion: Expanding the TelCovACT dataset to include more topics and platforms would provide a more realistic corpus for CT detection.
- Real-World Deployment: The researchers plan to collaborate with NGOs to optimize the real-world deployment of TelConGBERT, focusing on efficient data acquisition, user feedback mechanisms, and error rate communication.
- Complexity of CTs: The paper emphasizes that CT detection is inherently complex and often defies simple binary classification, highlighting the need for nuanced approaches.
Summary updated