Semi-supervised parser for semantic dependency parsing.

This repo contains the code used for the semi-supervised semantic dependency parser in Jia et al. (2020), Semi-Supervised Semantic Dependency Parsing Using CRF Autoencoders. Part of the codebase is extended from Parser-v3.

python3
tensorflow-gpu=1.13.1

Our semi-sdp parser can be trained by simply running

python3 main.py train UnlabelGraphParserNetwork  --force --config_file $CONFIGFILE

CONFIGFILE contains all hyperparameters of the model, we give an example file in the directory 'config/'. By default, if the save directory already exists, you'll get a prompt warning you that the system will delete it if you continue and giving you one last chance to opt-out. If you are debugging or want to run the program in the background, add the --force flag.
Our model was trained with Glove embedding.
To train with the unlabeled data, note the parameters in Flag subsection in the config file. Set fix_label_data=True under the Flag and the labeled_num means how many sentences have labels in your training set (Need to put labeled data in front of unlabeled data). For other parameters that need to be modified, please refer to the paper and code.

Format of datasets used in this code is CoNLL-U, we give an example in the directory 'data/'. The path of training/develop set is set in the CONFIGFILE, these can be changed according to yourself setting. Depending on the memory size (24G) of the running platform, we use labeled sentences less than 60 in length. Script that filters sentence according to lengths is simple, we give an exampel in the directory 'scripts/select_sents.py'.

The trained model can be run by calling

python main.py --save_dir $SAVEDIR run --output_dir TestResult $DATAFILE

This will save parsed sentences in DATAFILE to the TestResult/ directory--make sure no files in different directories have the same basename though, or one will get overwritten! The sentences in DATAFILE is in CoNLL-U format. SAVEDIR is the directory that your trained model saved.