In my experience, the problem you're trying to solve boils down to finding articles similar to a given search query that are in the given category. Trying to outsmart Lucene on that kind of a problem is going to be challenging given that it's for example used as a benchmark in research[1], so switching over to ElasticSearch is arguably the way to go.
There's a specific feature in Lucene called "MoreLikeThis", and it's also exposed in WP's search API to find articles similar to other articles. The documentation[2] of that feature provides a fairly good explanation of how it works, making it a possible starting point on how to filter a given document to improve the search results.
If I remember correctly there are a couple of research papers that study how to recommend sources for articles (or articles for a given source), but I'd have to go look for them to find them. You might want to consider searching the Research Newsletter archives and Google Scholar as that might give you a couple of existing approaches.
Footnotes: 1: A paper I reviewed for the Research Newsletter used it: https://meta.wikimedia.org/wiki/Research:Newsletter/2016/May#Evaluating_link... 2: https://lucene.apache.org/core/3_0_3/api/contrib-queries/org/apache/lucene/s...
Cheers, Morten
On 1 October 2017 at 18:36, Mukunda Modell mmodell@wikimedia.org wrote:
I think this is a really cool idea. I don't know of other similar tools but it does sound like something that should be a good fit for elasticsearch.
On Fri, Sep 29, 2017 at 9:34 AM Guilherme Gonçalves < guilherme.p.gonc@gmail.com> wrote:
Hi everyone,
I've been hacking on a new tool and I thought I'd share what (little) I have so far to get some comments and learn of related approaches from the community.
The basic idea would be to have a browser extension that tells the user if the current page they're viewing looks like a good reference for a Wikipedia article, for some whitelisted domains like news websites. This would hopefully prompt casual/opportunistic edits, especially for articles that may be overlooked normally.
As a proof of concept for a backend, I built a simple bag-of-words model of the TextExtracts of enwiki's Category:All_articles_needing_additional_references. I then set up a tool [1] to receive HTML input and retrieve the 5 most similar articles to that input. You can try it out in your browser [2], or on the command line [3]. The results could definitely be better, but having tried it on a few different articles over the past few days, I think there's some potential there.
I'd be interested in hearing your thoughts on this. Specifically:
- If such a backend/API were available, would you be interested in using
it for other tools? If so, what functionality would you expect from it?
- I'm thinking of just throwing away the above proof of concept and using
ElasticSearch, though I don't know a lot about it. Is anyone aware of a similar dataset that already exists there, by any chance? Or any reasons not to go that way?
- Any other comments on the overall idea or implementation?
Thanks!
1- https://github.com/eggpi/similarity 2- https://tools.wmflabs.org/similarity/ 3- Example: curl https://www.nytimes.com/2017/ 09/22/opinion/sunday/portugal-drug-decriminalization.html | curl -X POST http://tools.wmflabs.org/similarity/search --form "text=<-" -- Guilherme P. Gonçalves _______________________________________________ Cloud mailing list Cloud@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/cloud
Cloud mailing list Cloud@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/cloud