This is very interesting and definitely relevant to my interrogation.

Finance aside, it may be worth trying to define more in depth what would be a workflow, and usecases. For people to validate datas, they have to want it … Why would they want it ? It’s not enough to just build a random tool and hope it will be used. I find myself often testing tools proposed in the Wikidata Weekly, actually it rarely occurs that it become something I use very often.

We’re on a Wikimedia project, one of the purpose shared by people in it is to build an encyclopedy … One cool goal would be a tool that for example analyse the text produced by people and suggests wikilinks. People could then give feedback by saying to the tool « yes, this is indeed the meaning of this word in my text » to validate the output, or « yes, this is indeed a cool wikilink ». Just a thought.

On this example, those techs would be used to analyse the produced text and make suggestions. It could be made through an open API that receive (wiki)text, returns annotated wikitext, and recieve feedback from users through the client UI, eventually writing posiive feedback into Wikidata if it’s missing.

Such a service would be reusable for all kind of usecases. Another could be, for example, analysis of commons semi-structured descriptions to suggests values for depicts claims.

Le ven. 20 sept. 2019 à 14:22, Houcemeddine A. Turki <turkiabdelwaheb@hotmail.fr> a écrit :

Dear all,

I thank you for your efforts. To know more about word embedding and semantic similarity, please refer to the survey of our research group about the issue available at https://www.sciencedirect.com/science/article/pii/S0952197619301745. If you would like that we work on using these techniques to enrich Lexicographical Data on Wikidata, we will be honoured to do this. However, we will face two main problems. The first one is absolutely funding and the second one is that we need people to validate the information returned by these two techniques and adjust it if needed.

Yours Sincerely,

Houcemeddine Turki (he/him)

Medical Student, Faculty of Medicine of Sfax, University of Sfax, Tunisia

Undergraduate Researcher, UR12SP36

GLAM, Research and Education Coordinator, Wikimedia TN User Group

Member, Wiki Project Med

Member, WikiIndaba Steering Committee

Member, Wikimedia and Library User Group Steering Committee

Co-Founder, WikiLingua Maghreb

____________________

+21629499418

-------- Message d'origine --------
De : Thomas Douillard <thomas.douillard@gmail.com>
Date : 2019/09/20 12:08 (GMT+01:00)
À : "Discussion list for the Wikidata project." <wikidata@lists.wikimedia.org>
Objet : [Wikidata] Lexical datas and automated learning – where it is answered to « I don’t believe in Wikidata senses developpment »

I recently read the french sentence « Je ne crois pas au développement des sens. » — translation : I don’t believes senses with develop much (following links in a Wikidata Weekly summary, the slides on a french meeting about Wikidata lexicographical datas). I believe in it, (regardless of the arguments exposed in the slides), and I write this email to try to explain why.

I’m curious to know if there is already some work on the automated discovering of lexicographical datas / senses thanks to the help of Wikidata items.

There is tools for automated tagging of terms with the corresponding Wikidata item, that appeared on this mailing list and/or on the wikidata weekly summaries.

There is also methods that can discover senses into texts using only the terms with no reference to any external « sense » like https://towardsdatascience.com/word-embedding-with-word2vec-and-fasttext-a209c1d3e12c and can discriminate several usages of the same word according to the context.

Wikidata lexicographical datas and Wikibase items could close the loop between the 2 methods and allow us to semi automatically build tools that annotate texts with Wikidata items it there is something relevant in Wikidata, but if there is nono try to suggest to add datas on Wikidata, wether it’s a missing item or a missing sense for the term.

It may even be possible to store word embeddings generated by word2vec methods into Wikidata senses.

In conclusion, I think Wikidata senses will be used because they allow to close a gap. It does not depends only on a strong involvement in a volunteer traditional lexicographic community. If reasearchers of the language community dives into this and develop algorithms and easy to use tools to share there lexicographical datas in Wikidata, there could be a very positive feedback loop where numerous data ends to be added on Wikidata, where the store datas helps the algorithm to enrich text annotations, for example, and missing datas are semi automatically added thanks to user feedback.

This is all just wishful thinking, but I thought this deserved to be shared, hopefully this will launch at list a thread of ideas/comment in here :)

Thomas