Train a custom parser

we need to train a parser with custom semantics like this:
Can we do it with prodigy? i mean, can we add custom label for relations?

We need to process documents and get relathions between DATE, EVENT, LOCATION, PERSON
Our “root” is the EVENT and then we need to “link” to other entities and get relations. Is it possible to train a custom parser?

It will be very interesting to see the accuracy. The documents have many DATEs PERSONs etc so we must find a way to link them, is this the correct way for you?


You should be able to annotate the relations in Prodigy, yes. You won’t be able to use a model initially, but you can start with the EVENT annotated, and mark a relation to the child word. Once you’ve done that, you should have some simple trees, so you can switch to training spaCy. We don’t have a recipe for that yet, so you would need to export and transform the data.

You might also find it useful to start off with a rule-based approach, which uses the existing syntactic dependencies to suggest relations. Then you can mark these relations as accept or reject. If your rules have high recall, you may be able to simply train a model to replicate your binary decisions. This might be an easy way to extract some of your relations without a parsing-based approach.