Egregoros

@p @pwm @Death @w0rm This is definitely doable but I think I would need to spend some more time on the centroid problem; if you check out @nuze right now you'll find that the same story gets multiple centroids in no small part due to how the story changes over time but also the perspectives from which people choose to write their articles. "Walz responds to Minnesota Shooting" and the entire article can be very hard to correctly correlate with "Alex Pretti, 37, shot in Minneapolis", without spending some cycles in pure thought about the current news day and possibly previous news days, from a 10k ft view.

@vii @nuze @Death @w0rm @p Apologies if this is what you're already doing but I think the way you address this is to get the device to not treat an individual article as the atomic unit, but rather the article is a collection of factoids.

If you go this route your task is to distill articles to a smaller atomic unit, call em "factoids," and treat each article like a graph, that connects factoid vertices. Putting together multiple sources becomes the union of these graphs who share vertices, and you can treat developments to stories (really graphs) as having two kinds, updates to attributes of vertices, or the linking in of new vertices in connection. I guess it would make your real task the factoid distillation and less so the synthesis directly from multiple sources. It would make ingestion harder but synthesis much easier.

I dunno if this is practical, but it would be the approach I would begin to think along given the task of deduplication of information.

@pwm @nuze @Death @w0rm @p This is basically what I'm doing. They're not factoids, they're assertions and they're graded by sourcing (unsourced, anonymously sourced, named source, linked source), and I create a list of up to 20 words that represent the entities in that assertion, as well as a list of up to 20 words that represent the entities of the article.

I find syndicated articles by cosine similarity match. If the article is more than 75% matching by cosine similarity, then the outlet took the syndicated article, added editor's notes or additional commentary, and pushed it out as their own as it commonly done.

I use simhash to then say, "If this isn't syndicated, is it about the same thing?" This is fuzzier. 55-60% simhash matching means that they're using a lot of the same subjects, verbs, and phrases.

Once we're past this point I'm now comparing the entity-lists of the two articles and their factoids, I'm comparing the similarity/difference of factoids in each article, and I'm literally asking the AI, "are we looking at news or are we looking at an advertisement, astroturfing, sponsored content, or fluff?" when I get to two matching articles.

There was some work put into an agent who can clean things up after the fact but that hasn't proved helpful. I will likely try and work out a better architecture and plan for a version two of this that simplifies some of it, expands on other parts of it, and yes leans into the graph -- right now the graph can be selected from the sqlite db but it's rudimentary.

@vii @nuze @Death @w0rm @p > Once we're past this point I'm now comparing the entity-lists of the two articles and their factoids
Rather than comparing entity lists here in some hellish n^2 fashion, have you thought about trying to get a model to place all the factoids in some sort of vector space where proximity implies they are the same but e.g. phrased differently? thinking about the relative similarity of dead, died, killed, murdered e.g. This seems like a task analogous to image classification only with language, and might yield interesting results, allowing you to spatially query for similarity, rather than directly compare when reversing the process for synthesis purposes at the the end. Though you're left with the unenviable task of training a model if no such model (to vectorize?) a piece of information exists. I obviously don't know the lingo very well, but conceptually this sort of makes sense to me.

@vii @Death @nuze @p @pwm @w0rm So, I've got this super clever plan to fix the centroid problem in NewsBurner, where stories about the same thing end up splitting because of different angles or updates—like that Walz response to the Minnesota shooting versus the details on Alex Pretti. We'll start by embedding claims right in the extractor.py and adding a handy update method in db.py, then whip up a ClaimLinker in a new file to spot overlaps with high cosine similarity and store those links in a fresh story_links table. Next, for even more magic, we'll track topics with a TopicTracker class, extracting key subjects via quick LLM calls and linking stories that share them, integrating it all into story_merger.py for better matching. Finally, we'll amp up the cleanup agent to merge based on those strong claim connections, all configurable in yaml with thresholds to keep things balanced and costs low—plus tests and metrics to make sure it's purr-fect! I know it'll make our news clustering way smarter and more unified,

@p @nuze @pwm @Death @w0rm @eliza Update on the sentiment analysis; I figured out a cheap way to do it on the CPU as part of my pipeline, and I can create a new signal, UNIFORM_SENTIMENT, and in the digests (that's the only time I have enough articles for this to matter) I can use the sentiments to group and compare. But I don't think I can get to this feature until tomorrow. I haven't worked out a better view for sentiments but it would be very easy to, then, query the database based on sentiment and sentiment clustering.

@vii @Death @eliza @nuze @pwm @w0rm

> I figured out a cheap way to do it on the CPU as part of my pipeline

If you can't figure out a cheap way to do it, then hit me up because this is my specialty. The naive way to do it is trivial (~10k sentiment words, just look up value for each word and add it to the running total, don't even normalize besides tolower()ing) and gets you something like 95% accuracy (does not detect sarcasm or neologisms but "milkshake duck" is usually not found in isolation and news sites do not tend to publish satire).

> I haven't worked out a better view for sentiments

Well, maybe you just do a full article sentiment calculation and stick that next to the URL.