This is a reimplementation of the neural CCG supertagger by Lewis et al. (2016), using more modern methods.
The supertagger is trained on the lexical categories used in the CCGbank annotations. It uses XLM-RoBERTa encodings from Huggingface plus a simple linear layer to make the supertag predictions.
The top-1 tagging accuracy of this simple model is ~91% when trained on 02-21 and evaluated on 22.
You can instruct train.py to save the trained model in a file by specifying the filename in the parameter model_filename in config.yml.
You can use such a trained model to run the actual supertagger (supertagger.py). The supertagger has a number of useful command-line options, see the source code. It will generate a JSON file with supertag scores for the test corpus, which is suitable as input for the companion project ccg-parser.
Note that the supertagger suppresses CCGbank supertags that are not real CCG categories, such as ":" and ",", and removes words whose top-k supertags only contain such suppressed categories. This means that the sentence in the JSON file may not be exactly the same as in the original corpus, making a direct evaluation of parsing accuracy difficult.
During preprocessing, train.py caches a supertag lexicon in a file supertag_vocabulary.txt. This file is reused for later runs of train.py. If you want to retrain on a different dataset, delete this file first.
This Github repository already contains a file supertag_vocabulary.txt, which is the complete supertag vocabulary in Sections 02-21 of the CCGbank. Using this file will allow you to train a supertagger on just a subset of the CCGbank, which might otherwise not contain all the supertags you need for your test set. If you want to train a supertagger from scratch on a different treebank, you should delete this file first.