Luigi A.

Ok!

Here an example from my experiments - tell me a pathway between the concept of "life" and "artificial intelligence" ? These results are only obtained from a simplified knowledge network that does not know about types of entities: topics, people, place, etc. are all treated the same way. And yet, you see the reasoning is not trivial. I think adding a KG to guide the "reasoning" through types of topics and to unpack generic concepts (like "computer science") into backgrounds of people would really be ... excessively satisfying 🙂 (Apologise if I elaborated outside of the scope of bookclubontology)

I worked on multipartite graphs and I think that approach can be valid : recommender systems can be approximated to bipartite graphs (users, content), I aimed to to generalise recommendations between different types of entities, by leveraging on entities properties. As results, recommendations are good enough even if where there are no followers (or too few users). I would use graph analytics on top of KG, where KG may describe types of Q and A. Elaborating on your hint Matthias S. - Grouping of Personae could be done in different ways, right now thinking about similarity measures but in case of dense networks also dimensionality reduction may proof good to cluterize. I would groups types of questions and answers ( maybe bread for NLP teeth ), and ascribe those to types of "meta"-relationships in a KG, in attempt to abstract the scope of questions / answers from granularity of topics (like, imagine to map "raw ingredients" and "cooking details" into commodity classification and cooking procedures). And then use relationships and extracted topics as properties in topological analyses, to classify types of Persons. Inference could be done upon these results. I used random walks between entities; I think it could be done between subgraphs of persons-topics-persons. I tried random walks because reasoning does not always work well with shortest paths answers - but rather organising paths that are not trivial, and also not too long. They proved insightful. I also would like to explore new approaches. So to wrap up, I would use a KG to describe types of entities and relationships, and then use network science analytics to organise topics, persons, background, and then use statistical inference for suggestions of intermediate steps of learning paths. For the sake of knowledge - any public grant to apply for a #common work ? Or Robert M. does companies like McKinsey and similar may offer a small grant or would be receptive of a pilot proposal (cost-benefit opportunity 🙂 ) ?

Oh yeah inference of learning paths is big time. How would you deal with it matthias?

Imagine this. You have knowledge about nlp. You are interested in climate change. How d you get to apply nlp for cc ? In my opinion (my strategy would be) go backward. So cc may unpack in climate mitigation, sequestration etc. So an example may be structuring the onthologies of cc into kg, nlp may map the steps into a general human spoken question. I d find useful to see all the steps unpacked in a learning path (that s why i built my platforms). How granular are the steps ? In my opinion a step should be clear enough for a layperson, for generic use. I d also keep the logic distinct: why and how to navigate insights, other factual info( where what hetherogeneous data video media podcast etc) aggregated on top. It is important, in my opinion, what a user can do ... Wrapping up, i ll start from a domain (climate change, ethology and taxonomy, food discovery, bibliographic search....) and backward. Tech concepts are the last part and just functional to make the life of a passionated human being easier... Lol

So that public version does this : 1. Fetch context from a discovery engine 2. Map entities in queries to other data sources : articles (wikipedia), books (e.g. amazon and other sources), news (here, only geocalised for us ) and can support others (e.g. youtube or custom content or API s from an organisation )

My own discovery engine. Entities that exists in public knowledge graphs, like wikidata or wikipedia, are matched to pull in articles or other media relevant to that entity.