Re: Fact-checking and community notes on the Fediverse

Social Web Incubator Community Group,

Hello. I am pleased to share some preliminary brainstorming and ideas about decentralized fact-checking and argumentation using P2P filesharing networks.
Hopefully some of the following ideas can be of use for the Fediverse, e.g., for the discovery of existing annotations.

Introduction
With respect to sharing Web Annotations, uses of P2P networks have been previously explored (Segawa, 2006). Providing users with access to these kinds of networks from their Web browsers, today, is possible with WebRTC (Werner & Vogt, 2014; Ersson & Siri, 2015).
P2P filesharing networks could be of use for decentralized fact-checking and argumentation. Facts or claims could be stored in entries, a special kind of file resource.
By creating and sharing digitally-signed user feedback, notes, comments, or annotations with respect to those facts or claims in entries, users could express their determinations with respect to the veracity of facts or claims and could also present arguments for or against them (Bex, Snaith, Lawrence, & Reed, 2014).
Entries could contain one or more references to paraphrases of content from locations on the Fediverse (see: Appendix A). Annotation objects from the Fediverse could be indexed and redundantly stored on P2P filesharing networks.
Uses of Embedding Vectors
Instead of, or in addition to, using cryptographic hashes to index and address content on P2P networks, digitally-signed entries for facts or claims could be indexed and addressed using embedding vectors (Zaarour & Curry, 2022).
As considered, entries would be a special kind of file resource where their embedding vectors, embedding vectors verifiably for selections of other resources' contents, would be stored inside of them (see: Appendix A) rather than obtained from processing them with AI models.
Indexing and addressing entries thusly would allow them to be merged or wrapped, e.g., to add paraphrases, digitally signing them at each step, without having to reindex them. Modifications, however, would result in changes to entries' cryptographic hashes.
Deep learning can be used to detect and identify sentential paraphrases (Zhou, Qiu, Liang, & Acuna, 2022). More elaborate uses of language models could be utilized for inquiring and reasoning about whether sentences occurring in contexts were paraphrases.
With respect to fact-checking on the Web, scenarios to consider include both fact-checking content which was expressly indicated to be a fact or claim by their authors, e.g., using custom elements, and fact-checking arbitrary selections of documents' content.
Explorations with respect to fact-checking arbitrary selections of content include the open-source Citation Needed project by the Future Audiences team of the Wikimedia Foundation.
The Prompt API
Exploration is underway into providing APIs for accessing language models in Web browsers; the Web Machine Learning Working Group is developing the Prompt API.
With access to language models in Web browsers, users might be able to obtain embedding vectors for portions of content in Web documents. These embedding vectors could be used to search for other content, e.g., annotations, including on P2P networks.
Custom Elements
HTML5 custom elements could allow facts or claims to be expressed in documents, e.g., to add visual indictors near them or enable special context menus for them, while specifying values for embedding vectors computed for them using AI models (see: Appendix C).
Appendices
Appendix A shows a markup sketch for an entry, a created entry wrapped to add a paraphrase to it.
Appendix B shows that embedding vectors could be added to Magnet URIs and Metalinks.
Appendix C shows that HTML5 custom elements could be used for asserted facts or claims which refer to entries on P2P networks by means of one or more embedding vectors.
Appendix D shows an approach involving shortcodes for authors using content-management systems to be able to easily add facts or claims to their content.
Bibliography
Bex, Floris, Mark Snaith, John Lawrence, and Chris Reed. "ArguBlogging: An application for the argument web." Journal of Web Semantics 25 (2014): 9-15. https://www.sciencedirect.com/science/article/pii/S1570826814000079
Ersson, Kerstin, and Persson Siri. "Peer-to-peer distribution of web content using WebRTC within a web browser." (2015). https://www.diva-portal.org/smash/get/diva2:819420/FULLTEXT01.pdf
Segawa, Osamu. "Web annotation sharing using P2P." In Proceedings of the 15th international conference on World Wide Web, pp. 851-852. 2006. http://ra.ethz.ch/CDstore/www2006/devel-www2006.ecs.soton.ac.uk/programme/files/pdf/p45.pdf
Werner, Max Jonas, and Christian Vogt. "Implementation of a browser-based P2P network using WebRTC." Hamburg (2014). https://inet.haw-hamburg.de/teaching/ws-2013-14/master-project/Prj1-report-werner-vogt.pdf
Zaarour, Tarek, and Edward Curry. "SemanticPeer: A distributional semantic peer-to-peer lookup protocol for large content spaces at internet-scale." Future Generation Computer Systems 132 (2022): 239-253. https://www.sciencedirect.com/science/article/pii/S0167739X22000590
Zhou, Chao, Cheng Qiu, Lizhen Liang, and Daniel E. Acuna. "Paraphrase identification with deep learning: A review of datasets and methods." arXiv preprint arXiv:2212.06933 (2022). https://arxiv.org/pdf/2212.06933


Appendix A: Sketch of an Entry for a Fact or Claim

<action kind="add-paraphrase">

  <base>

    <action kind="create">

      <base />

      <time>2024-01-14T00:01:00Z</time>

      <v id="v-1" model=" urn:ai:model:llama:3.2:90B">...</v>

      <metalink id="source-1">

        <file name="article1.html">

          <url>https://www.example1.com/user1/article1.html</url>

        </file>

      </metalink>

      <selection source="source-1">

        ... <select v="v-1">A sentence.</select> ...

      </selection>

      <signature>...</signature>

    </action>

  </base>

  <time>2024-01-14T00:00:00Z</time>

  <v id="v-2" model="urn:ai:model:llama:3.3:70B">...</v>

  <metalink id="source-2">

    <file name="article2.html">

      <url>https://www.example2.com/user2/article2.html</url>

    </file>

  </metalink>

  <selection source="source-2">

    ... <select v="v-1 v-2">A paraphrase.</select> ...

  </selection>

  <signature>...</signature>

</action>


Appendix B: Adding Embedding Vectors to Magnet URIs and Metalinks
Embedding vectors could be added to Magnet URIs by means of adding a key: xv.
Embedding vectors could be new components of metalinks.
<metalink xmlns="urn:ietf:params:xml:ns:metalink">
  <published>2009-05-15T12:23:23Z</published>
  <file name="example.txt">
    <url>http://www.example.com/example.txt</url>
    <vector model="urn:ai:model:llama:3.3:70B">...</vector>
  </file>
</metalink>

Appendix C: Custom Elements for Facts or Claims
A custom element could be used to signify an asserted fact or claim, referring to an entry on a P2P network by means of embedding vectors alongside other information. Via a JavaScript library, and perhaps WebRTC, clients could participate in P2P networks and retrieve entries, feedback on entries, or both.
Notice that, for the special file type of entries, those embedding vectors within them and not of the XML file itself are utilized with respect to storing and addressing the resource on P2P networks.
<verifiable-claim see="magnet:?xv=...">Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.</verifiable-claim>
Appendix D: Content Authoring with Shortcodes
How might authors easily add facts or claims to their content? With respect to popular content-management systems, the syntax for so doing could resemble that of existing shortcodes like [quote].
[claim]Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.[/claim]
During content-publishing processes, authors' content-management systems (e.g., Drupal, WordPress) – or configurable plugins or extensions for these systems – could handle searching for existing paraphrases, adding new facts or claims (if needed) to P2P filesharing networks, obtaining the data for use in the see attributes, caching these data, and generating markup.

________________________________
From: Emelia S. <emelia@brandedcode.com>
Sent: Monday, January 13, 2025 11:21 AM
To: Evan Prodromou <evan@prodromou.name>
Cc: public-swicg@w3c.org <public-swicg@w3c.org>
Subject: Re: Fact-checking and community notes on the Fediverse

This is already something on the list of things that the ActivityPub Trust  & Safety Taskforce is working on:

<https://github.com/swicg/activitypub-trust-and-safety/issues/4>
[4.png]
Idea: Annotations / Labeling of content · Issue #4 · swicg/activitypub-trust-and-safety<https://github.com/swicg/activitypub-trust-and-safety/issues/4>
github.com<https://github.com/swicg/activitypub-trust-and-safety/issues/4>

The Web Annotations model could work, but the discovery of annotations that exist is the hardest part, I've started solving that in https://github.com/ThisIsMissEm/annotations-service where I use the sha256 hash of the Object ID as the annotation collection ID, giving a very simple way to fetch all annotations for a given object.

I do want to investigate what an Annotate activity would look like, but I suspect this would just be an announcement of sorts "hey, there's this web annotation over here for this target"

Yours,
Emelia

On 13 Jan 2025, at 04:23, Evan Prodromou <evan@prodromou.name> wrote:

We don't have an easy way for remote actors to annotate content on the Fediverse.

The biggest use case for this is to have permissionless fact-checking or community notes. A fact-checking service could annotate a remote content object like a Note or a Video with additional fact-checking information, and compliant clients or servers could show the fact-checking information when showing the Note.

I think there are some tricky parts to this structure, which I believe suggests that we should start working on it.

Evan