Indeed, the purpose does matter. Is the end goal the content similarity of articles themselves (perhaps say to detect articles that might be merged) or is the end goal the relatedness of topics represented by those articles? If the latter is the goal, then the Wikipedia category system relates articles with some commonality of topic, and distance between articles via the category hierarchy is an indicator of levels of relatedness. Similarly navboxes relate articles that have something in common, as do list articles. All of these three things are manually curated, and may be a much cheaper way to determine relatedness of topics than messing about with bags of words, etc. But it all really depends on the end goal. Kerry -----Original Message----- From: Wiki-research-l [mailto:wiki-research-l-bounces@lists.wikimedia.org] On Behalf Of Isaac Johnson Sent: Wednesday, 8 May 2019 1:35 AM To: Research into Wikimedia content and communities <wiki-research-l(a)lists.wikimedia.org> Subject: Re: [Wiki-research-l] Content similarity between two Wikipedia articles Hey Haifeng, On top of all the excellent answers provided, I'd also add that the answer to your question depends on what you want to use the similarity scores for. For some insight into what it might mean to make choose one approach over another, see this recent publication: https://dl.acm.org/citation.cfm?id=3213769 At a high level, I'd say that there are three ways you might approach article similarity on Wikipedia: * Reader similarity: two articles are similar if the same people who read one also frequently read the other. Navigation embeddings that implement this definition based on page views were generated last in 2017, so newer articles will not be represented, but here is the dataset [ https://figshare.com/articles/Wikipedia_Vectors/3146878 ] and meta page [ https://meta.wikimedia.org/wiki/Research:Wikipedia_Navigation_Vectors ]. The clickstream dataset [ https://dumps.wikimedia.org/other/clickstream/readme.html ], which is more recent, might be used in a similar way. * Content similarity: two articles are similar if they contain similar content -- i.e. in most cases, similar text. This covers most of the suggestions provided to you in this email chain. Some are simpler but are language-specific unless you make substantial modifications (e.g., ESA, the LDA model described here: https://cs.stanford.edu/people/jure/pubs/wikipedia-www17.pdf) while others are more complicated but work across multiple languages (e.g., recent WSDM paper: https://twitter.com/cervisiarius/status/1115510356976242688). * Link similarity: two articles are similar if they link to similar articles. Generally, this approach involves creating a graph of Wikipedia's link structure and then using an approach such as node2vec to reduce the graph to article embeddings. I know less about the current approaches in this space, but some searching should turn up a variety of approaches -- e.g., Milne and Witten's 2008 approach [ http://www.aaai.org/Papers/Workshops/2008/WS-08-15/WS08-15-005.pdf ], which is implemented in WikiBrain as Morten mentioned. There are also other, more structured approaches like ORES drafttopic, which predicts which topics (based on WikiProjects) are most likely to apply to a given English Wikipedia article: https://www.mediawiki.org/wiki/Talk:ORES/Draft_topic On Tue, May 7, 2019 at 9:54 AM <fn(a)imm.dtu.dk> wrote: -- Isaac Johnson -- Research Scientist -- Wikimedia Foundation _______________________________________________ Wiki-research-l mailing list Wiki-research-l(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiki-research-l