using custom NER

[vlordier]

Could you point out how you would go about using a custom trained NER, with domain-specific entities ?

[Mischn]

Hello Vincent,

please correct me if this does not answer your question, but I assume you want to know how one can extract named entities as relational arguments.

In a nutshell, the approach of Graphene is based on a prior sentence-simplification process, where really complex sentences are recursively transformed into simpler ones. After that, subject-predicate-object relationships are extracted from these simpler sentences (for more details see the Graphene publication).

In this way, our product acts as preprocessing framework for any system that is capable of extracting subject-predicate-object relationships from texts. So, in order to extract (domain-specific) NEs as arguments, you can simply implement your own RelationExtractor and use it as the RelationExtraction component inside of Graphene.

Here is what you will have to do:

• You need an RelationExtractor that will identify named entities (using NER) and their relationships in a sentence. You can implement your own Extractor or modify one of the already available implementations in the package org.lambda3.graphene.core.relation_extraction.impl
• Ensure that your Implementation inherits from the class org.lambda3.graphene.core.relation_extraction.RelationExtractor
• To use your implementation in the framework, you need to specify your new implementation in the configuration file (reference.conf). Simply change the attribute:

relation-extractor = org.lambda3.graphene.core.relation_extraction.impl.HeadRelationExtractor to point to your new RelationExtractor-class: relation-extractor = org.lambda3.graphene.core.relation_extraction.impl.MyNERRelationExtractor

[vlordier]

All clear Matthias ! Thank you :)