search

dice-group / FOX

Federated Knowledge Extraction Framework
GNU Affero General Public License v3.0
188 stars 52 forks source link

Question: Import a Knowledge Base for Entity Extraction #53

Closed ali1k closed 5 years ago

ali1k commented 5 years ago

Is it possible to use our custom knowledge graphs/ontologies/thesaurus within the FOX? If yes, is there any documentations available about that?

renespeck commented 5 years ago

Hi, thanks for the question.
The linking comes with the help of the AGDISTIS/MAG web service. To integrate another KG you need to integrate it in the AGDISTIS/MAG web service and update the endpoint of the service in the FOX config files.

I hope that helps.

ali1k commented 5 years ago

Thanks for the answer. I get that for linking and disambiguation you use AGDISTIS. However, I was mainly thinking about custom word embeddings provided by a knowledge base other than DBpedia.

DiegoMoussallem commented 5 years ago

@ali1k can you explain it a bit more how to custom word embeddings? for example, you can use AGDISTIS/MAG with Wikidata and other KGs.

ali1k commented 5 years ago

What I mean is to spot entities based on indexes coming from a certain knowledge graph. I am looking for a custom NER tool that that can be adapted to different domains.

RicardoUsbeck commented 5 years ago

rasa nlu provides a quick and dirty solution @vdanielupb can help

ali1k commented 5 years ago

thanks for the link, will check https://rasa.com/docs/nlu but would be great if @vdanielupb can give me more specifications on how to customise that.

vdanielupb commented 5 years ago

@ali1k To learn a model for entity extraction with Rasa NLU you can use the following steps:

  1. Generate a training file as described here: https://rasa.com/docs/nlu/dataformat/
  2. Write a configuration file for Rasa NLU https://rasa.com/docs/nlu/components/, so that it learns a CRF for entity extraction.
  3. Use the Python API https://rasa.com/docs/nlu/python/ of Rasa NLU to train the model
ali1k commented 5 years ago

@vdanielupb thanks, that is already very helpful.