Inverting the Web
We use search engines because the Web does not support accessing documents by anything other than URL. This puts a huge amount of control in the hands of the search engine company and those who control the DNS hierarchy.
Given that search engine companies can barely keep up with the constant barrage of attacks, commonly known as "SEO". intended to lower the quality of their results, a distributed inverted index seems like it would be impossible to build.
@freakazoid Shifting ground (and jumping back up this stack -- we've sorted the URL/URI bit):
What you suggest that's interesting to me is the notion of _self-description_ or _self-identity_ as an inherent document characteristic.
(Where a "document" is any fixed bag'o'bits: text, audio, image, video, data, code, binary, etc.)
Not metadata (name, path, URI).
*Maybe* a hash, though that's fragile.
What is _constant_ across formats?
@freakazoid So, for example:
I find a scanned-in book at the Internet Archive, I re-type the document myself (probably with typos) to create a Markdown source, and then generate PDF, ePub, and HTML formats.
What's the constant across these?
How could I, preferably programmatically, identify these as being the same, or at least, highly-related, documents?
MD5 / SHA-512 checksums will identify _files_, but not _relations between them_.
Can those relations be internalised intrinsically?
@freakazoid Or do you always have to maintain some external correspondence index which tells you that SOURCE.PDF was the basis for RETYPED.MD which then generated RETYPED.MD.ePub and RETYPED.MD.html, etc.
Something that will work across printed, re-typed, error/noise, whitespace variants. Maybe translations or worse.
Word vectors? A Makefile audit? Merkel trees, somehow?
@dredmorbius We have real world solutions for these problems in the form of notaries, court clerks, etc. I.e. (registered) witnesses. Trusted third parties, but they don't have to be a single party.
@dredmorbius In the RDF world I guess one doesn't sign the individual triple but the entire graph.
And it might make more sense to call these 4-tuples, because it's really "this person says that this object is related in this way to this other object".
@freakazoid So for 4-tuple:
4. Obect1-Object-2 relation
"Signed" means that the whole statement is then cryptographically signed, making it an authenticatable statement?
@freakazoid And, so:
Back to search and Web:
- The actual URL and path matter to the browser.
- They may matter to me. Some RoboSpam site ripping off my blog posts _might_ leave the content unchanged, but they're still scamming web traffic, ads revenue, or reputation, based on false pretences. I want to read my content from my blog, not SpamSite, even if text and hashes match.
@dredmorbius FWIW word and phrase presence/frequency is self-description, in that it is verifiable without consulting a human. It's also useful for search, though it's generally not what humans care about directly even though it's what they search on; what they care about is the actual idea or thing they think documents having those words or phrases might be about.
I need to check on what state-of-the-art is, but based on tuples or ngrams of even short word sets (2-3, maybe 4), you can create an extensive signature of a text sampling within it. You can transform those to be constant against various modulations (e.g., ASCII7 vs. Unicode, whitespace, punctuation, ligatures, even common spelling variants/errors).
And then check an offered text against a known signature on a sampling of tuples through the doc.
This undoubtedly exists.
@freakazoid AFAIU that's basically how plagiarism-detection works.
@dredmorbius They use techniques like this to detect plagiarism. You can compute something like a Bloom filter for a document and then use Hamming distance to compare. That can work well as long as one is not intentionally trying to defeat it.
Of course, that assumes raw text. Once you get into complex markup, the markup can change the meaning of the document without changing what a text extractor will see. And then there's higher-bandwidth media like images, audio, and viceo.
@freakazoid And for anyone following this:
I'm not an expert, though I'm interested in the area.
I feel like I'm staggering drunk in the dark. Some of what I'm describing is Things I Have Known for Five Minutes Longer Than You (or a few days). Some longer.
This is ... remote from most work I've done, though I've been kicking around ideas for a few years, and know at least _some_ of what I'm talking about.
Informed input / corrections welcomed.
Everyone is welcome as long as you follow our code of conduct! Thank you. Mastodon.cloud is maintained by Sujitech, LLC.