Model based on HuggingFace T5

[bjascob]

Just FYI in case someone is interested...

I uploaded to amrlib a parse model based on the pre-trained HuggingFace T5-base model. This scores an 81 smatch on LDC2020T02. I can't say I've put much work into optimizing params (ie... # of epochs / choosing the best epoch) and there's also the T5-large model to try, so if someone wants to try to push the SoTA with these, the pretrained transformers are a good place to start.

The format for the graph serialization (model input) can impact the results. Originally, I simply removed all the variables and saw an 82 smatch score but noticed that graphs with multiple nodes of the same name (ie.. 2 different "people" nodes) were merged so I added an _xx to the names to make them unique. This actually reduces the smatch score a bit; I assume because the transformer has trouble with the numbering, but at least it's trying not to merge the nodes. I wasn't quite clear from your paper if you are doing something similar to this or just allowing multiple nodes in the same graph to be represented by the same string.

I also looked at the RikVN scripts that serialize/de-serialize the graphs to see if they had a better way to handle this. One thing to note about these scripts if they appear to be lossy. If you simply serialize / de-serialize and skip the model completely you get about a 0.98 smatch score, meaning there is some small error introduced by this. The serialize/de-serializer I'm using gives a 1.0 smatch so it shouldn't be introducing any error, though I can't be sure it's the "best" possible serialization format.

I'll be interested to hear if someone improves upon these numbers. It seems to me there's at least a few things to try that might push the results up a couple of points.