Generating Media Background Checks for Automated Source Critical Reasoning

TL;DR

The paper introduces Media Background Checks (MBCs) to capture signals of source trustworthiness and tendency for source-critical reasoning in NLP. It builds a 6,709-item MB/FC-derived dataset and demonstrates that retrieval-augmented generation improves MBC quality, with open-source models performing competitively. Human evaluations show MBCs aid both humans and NLP models in reasoning about untrustworthy evidence, increasing information sufficiency and reducing cognitive load. The work outlines a practical, interpretable path toward source-aware reasoning and releases data and code to advance research in this area.

Abstract

Not everything on the internet is true. This unfortunate fact requires both humans and models to perform complex reasoning about credibility when working with retrieved information. In NLP, this problem has seen little attention. Indeed, retrieval-augmented models are not typically expected to distrust retrieved documents. Human experts overcome the challenge by gathering signals about the context, reliability, and tendency of source documents - that is, they perform source criticism. We propose a novel NLP task focused on finding and summarising such signals. We introduce a new dataset of 6,709 "media background checks" derived from Media Bias / Fact Check, a volunteer-run website documenting media bias. We test open-source and closed-source LLM baselines with and without retrieval on this dataset, finding that retrieval greatly improves performance. We furthermore carry out human evaluation, demonstrating that 1) media background checks are helpful for humans, and 2) media background checks are helpful for retrieval-augmented models.

Figures (16)

• Figure 1: Retrieval-augmented NLP models can inadvertently misinform users if uncritically relying on retrieved documents from such untrustworthy sources. In preliminary experiments we found evidence of this occurring in practise: One popular search-augmented chatbot engaged in war-crimes denial after relying on Syrian state news to answer questions.
• Figure 2: We propose to generate Media Background Checks (MBCs) that summarise indicators of trustworthiness and tendency. MBCs can be used, either by humans or by retrieval-augmented models, to determine which documents can be relied on for further reasoning, and to craft reliable narratives based on untrustworthy evidence.
• Figure 3: Example background checks for Natural News. The gold example is taken from the Media Bias / Fact Check website, while the generated example is produced by GPT-3.5 augmented with Google search as described in Section \ref{['section:models']}. The gold example has been shortened, and the full version can be seen at https://mediabiasfactcheck.com/natural-news/.
• Figure 4: Prompt used for ChatGPT and Llama 3 when generating MBCs with no supporting retrieved evidence. This prompt is also used to generate the initial MBC which is later updated in the retrieved-evidence setting.
• Figure 5: Prompt used for ChatGPT and Llama 3 when updating an MBC with retrieved information. The retrieved information is input to the prompt in the form of question-answer pairs, following the methodology described in Section \ref{['section:models']}.