Hacker News Reader

Comment by wolverine876 | Hacker News Reader

I'd love a RSS reader that groups items on the same topic.

For example, let's say I have feeds from 25 news sources and the President gives a speech on Ukraine. I end up with 25 articles on it, spread across my item list; I need to process it 25 times. IME, most items are duplicated at least once - half my items are not needed.

If the feed reader would group all the Biden speech articles then I could process it once - pick the item I want to read, delete the rest. That's also more efficient because it's easier to compare the items.

Online news aggregators like Google News already do that, so the technology exists, but I haven't seen it in a feed reader. I also wonder if the LLMs could group them more accurately.

View on HN

>> I'd love a RSS reader that groups items on the same topic.

We had the same concern with everyday industry news, and so did some of our customers, so we built it a while back!

Our Newsroom feed reader uses NLP and ML to group similar articles for low-noise research/reading:

https://newsloth.com/#features

Currently, it detects and groups articles from a few hundred popular global news sites, and gradually growing to cover more sources.

We also have a feed builder/generator that turns unstructured public webpages into structured RSS feeds, to make it a flexible and unified feeds platform:

https://newsloth.com/product/feed-builder

That's great! I will take a look.

I've been using Liferea [1] for years, and while it doesn't do automated content analysis, it does allow you to group feeds into folders, then get an aggregated view of the folder as a whole. It also allows "search folders", which are saved cross-feed searches that are exposed through the same folder interface.

I use Liferea as my desktop client, but with TT-RSS running on my VPS as the underlying data source.

[1]: https://lzone.de/liferea/ https://github.com/lwindolf/liferea

I used to use an RSS reader that had "smart folders" (filters) so I could group cross feeds and across servers but the author seems to be too busy to keep it going.

I bet language models could definitely help here, yeah. Perhaps something like (1) get content offeed items as they come in, (2) embed the content, (3) use those embeddings to group items. Probably not that difficult to be honest

Yep. That would be a classic sort of k-means problem. Just throw them all into a standard embedding, like the OA API embeddings, run k-means from sci-kit, then convert them into a list-of-lists: one RSS item (containing a list of title-URLs) per cluster.

The problem with this approach is determining the what k is for the k-means. But again, we could use the “elbow” technique to determine what’s the optimal k and then start grouping them together. I wonder if there are any automatic sophisticated clustering algorithms?

Hierarchical and DBSCAN don’t require upfront knowledge about the number of clusters.

Reply on news.ycombinator.com