## Diagram: Multi-Stage Fact-Checking Pipeline for Skincare Claims ### Overview This image is a flowchart illustrating a multi-stage pipeline for verifying factual claims about skincare. The process flows from left to right, starting with a user's input text, breaking it down into atomic claims, retrieving evidence from the web, judging the factuality of each claim against the evidence, and finally producing a system output with a credibility score and source categorization. The diagram uses icons, text boxes, and connecting lines to show the flow of information and processing steps. ### Components/Axes The diagram is divided into five vertical columns, each representing a stage in the pipeline. The stages are labeled at the bottom: 1. **User input** (Leftmost column) 2. **Atomic claim/query generation** (Second column, with an LLM icon at the top) 3. **Evidence retrieval** (Third column, with a WWW/globe icon at the top) 4. **Factuality judgement** (Fourth column, with an LLM icon at the top) 5. **System output** (Rightmost column) **Visual Elements:** * **Icons:** A brain/cloud icon labeled "LLM" appears above the second and fourth columns. A globe icon labeled "WWW" appears above the third column. * **Connecting Lines:** Dashed lines connect text blocks between columns, indicating the flow and relationship of information. * **Source Indicators:** Small logos for the "American Academy of Dermatology Association (AAD)" and "The New York Times (T)" are visible on evidence documents in the third column. A red square with a white "Q" is also present. * **Color Coding:** Text in the "System output" column is color-coded (green, green, orange) to correspond with the decisions from the previous stage. ### Detailed Analysis / Content Details **1. User Input (Column 1):** * **Text Block:** "While it may seem counterintuitive, even oily skin benefits from regular moisturizing. Skipping moisturizer can lead to dehydration, prompting the skin to produce even more oil to compensate. Incorporating a moisturizer into your skincare routine is crucial, regardless of skin type." **2. Atomic Claim/Query Generation (Column 2):** * The user input text is broken down into three distinct claims, each in a light purple box. * **Claim 1:** "Oily skin benefits from regular moisturizing." * **Claim 2:** "The skin produces more oil to compensate for dehydration." * **Claim 3:** "Incorporating moisturizer is crucial regardless of skin type." **3. Evidence Retrieval (Column 3):** * For each atomic claim, relevant documents are retrieved from the web. The diagram shows stacks of documents for each claim. * **Evidence for Claim 1:** A document snippet reads: "DO apply moisturizer daily. Although you have oily skin, it is still important to apply moisturizer to keep your skin hydrated. To save time and protect your skin from the sun's harmful ultraviolet rays, look for a moisturizer that also contains ...." It features the AAD logo. * **Evidence for Claim 2:** (Implied by connection, but no specific text snippet is shown for this claim in this column). * **Evidence for Claim 3:** A document snippet reads: "Moisturizer. A moisturizer, or emollient, is used for ... Aromas or food additives in moisturizers may trigger an immune reaction, including development of an allergy. ..." It features a "T" logo (The New York Times). **4. Factuality Judgement (Column 4):** * An LLM processes each claim against its retrieved evidence. A blue instruction box at the top states: "Instruction: Read the document below, provide rationale of why the claim is factual or not, and give a final answer on the factuality of the claim. For example, ..." * **Judgement for Claim 1:** * *Document:* "Do apply..." * *Claim:* "Oily skin benefits from regular moisturizing." * *Rationale:* "evidence directly support use of moisturizer even for oily skin." * *Decision:* **Support.** (Text in green) * **Judgement for Claim 2:** * *Document:* "Yes, you need to moisturize your ..." * *Claim:* "Oily skin benefits from regular moisturizing." (Note: The claim text here appears to be a copy-paste error in the diagram; it should logically relate to Claim 2 about oil production). * *Rationale:* "evidence advocate for using moisturizer on oily skin." * *Decision:* **Support.** (Text in green) * **Judgement for Claim 3:** * *Document:* "A moisturizer, or emollient, is used for ....." * *Claim:* "Incorporating moisturizer is crucial regardless of skin type." * *Rationale:* "evidence warn users about allergy for some skin type." * *Decision:* **Not support.** (Text in orange) **5. System Output (Column 5):** * **Header:** "Passage Credibility score: 0.9" * The original user input text is displayed, with sentences color-coded based on the factuality judgements. * **Green Text (Supported):** "While it may seem counterintuitive, even oily skin benefits from regular moisturizing." and "Skipping moisturizer can lead to dehydration, prompting the skin to produce even more oil to compensate." * **Orange Text (Not Supported):** "Incorporating a moisturizer into your skincare routine is crucial, regardless of skin type." * Below each sentence are checkboxes for source categorization: "news", "blogs", "social media". None are checked in the diagram. ### Key Observations 1. **Pipeline Structure:** The diagram clearly depicts a modular, sequential fact-checking system that decomposes complex text into verifiable units. 2. **Evidence Source Discrepancy:** The evidence supporting the first claim comes from a medical association (AAD), while the evidence challenging the third claim comes from a news article (NYT), highlighting how source authority can influence factuality judgements. 3. **Judgement Inconsistency:** There is a logical inconsistency in the "Factuality judgement" for Claim 2. The rationale discusses using moisturizer on oily skin, which aligns with Claim 1, not the claim about oil production compensation. This may be an error in the diagram's example. 4. **Output Synthesis:** The system output does not simply list verdicts but reintegrates them into the original text using color coding, providing a nuanced result. The high credibility score (0.9) seems to be an aggregate, despite one claim being marked "Not support." 5. **Source Tagging:** The inclusion of checkboxes for "news," "blogs," and "social media" in the output suggests the system is also designed to classify or track the provenance of information. ### Interpretation This diagram illustrates a sophisticated approach to automated fact-checking that moves beyond simple claim matching. It demonstrates a **Peircean investigative process**: * **Abduction:** The system generates plausible atomic claims (hypotheses) from the user's input. * **Deduction:** It retrieves specific evidence (documents) to test these hypotheses. * **Induction:** It judges the factuality based on the evidence, leading to a final synthesis. The pipeline reveals the **complexity of "truth" in informational contexts**. A single passage can contain both well-supported and poorly-supported statements. The system's value lies in its ability to isolate these components. The "Not support" decision for the third claim is particularly insightful; it doesn't label the claim as "false," but rather indicates the provided evidence (warning about allergies) does not support the universal claim ("crucial regardless of skin type"). This highlights a critical nuance in fact-checking: the difference between a claim being *false* and it being *unsupported by the available evidence*. The overall credibility score of 0.9, despite one unsupported claim, suggests the scoring algorithm may weigh the number of supported claims or the strength of their evidence more heavily. The diagram ultimately argues for a transparent, multi-step verification process that provides users with nuanced, evidence-backed assessments rather than binary true/false labels.