Hi all, I just checked in a usable proof-of-concept for a neural (RoBERTa-based to be specific) negation classifier. The way it works is a tiny bit of python code (using FastAPI) sets up a REST interface that runs the classifier: ctakes-assertion/src/main/python/negation_rest.py
it runs a default model that I trained and uploaded into Huggingface modelhub. It will automatically download the first time the server is run. there is a startup script there too: ctakes-assertion/src/main/python/start_negation_rest.sh The idea would be to run this on whatever machine you have with the appropriate GPU resources and it creates 3 REST endpoints: /negation/initialize -- to load the model (takes longer the first time as it will download) /negation/process -- to classify the data and return negation values /negation/collection_process_complete -- to unload the model to mirror UIMA workflows. Then, the UIMA analysis engine sits in: ctakes-assertion/src/main/java/org/apache/ctakes/assertion/ae/PolarityBertRestAnnotator.java The main work here is converting the cTAKES entities/events into a simpler data structure that gets sent to the python REST server, making the REST call, and then converting the classifier output into the polarity property. Performance: The accuracy of this classifier is much better in my testing. I am looking forward to being able to hopefully make the path to improving the performance easier as it can potentially just be a change to the model string to have it grab a new model on modelhub. The speed is marginally slower if we do a 1-for-1 swap, but that's a little bit misleading, because we currently run 2 parsers to generate features for the default ML negation module. If we don't need those parsers we can dramatically cut the speed of the processing even with the neural negation module. I tested this with the python code running on a machine with a 1070ti. The goal for these methods going forward if we want to scale should be to have the neural call do a few things with a single pass, especially if we are using large transformer models. But this proof of concept of a single task will hopefully make it easier for other folks to do that if they wish. FYI, another way of doing this is by using python libraries like cassis and actually having python functions be essentially UIMA AEs -- I think there will be a place for both approaches and I'm not trying to wall off work in that direction. Tim
