UCCA Parser

An implementation of “HLT@SUDA at SemEval 2019 Task 1: UCCA Graph Parsing as Constituent Tree Parsing“.

This version of the implementation uses lexical features in the corpus, including POS tags, dependency labels and entity labels, just as described in the paper.

For other models or versions, please see different branches.

Requirements

python >= 3.6.0
pytorch == 1.0.0
ucca == 1.0.127

Note that the code has not been tested on the newest version of ucca module.

Datasets

The datasets are all provided by SemEval-2019 Task 1: Cross-lingual Semantic Parsing with UCCA. The official website is https://competitions.codalab.org/competitions/19160.

Pre-trained embeddings: http://fasttext.cc

Performance

Here are the results I re-ran on June 13, 2019, which are almost the same as the results in the paper.

Usage

You can start the training, evaluation and prediction process by using subcommands registered in parser.cmds or just use the shell scripts included in.

$ python run.py -h
usage: run.py [-h] {train,predict,evaluate} ...
UCCA Parser.
optional arguments:
  -h, --help            show this help message and exit
Commands:
  {train,predict,evaluate}
    train               Train a model.
    predict             Use a trained model to make predictions.
    evaluate            Evaluate the specified model and dataset.

Optional arguments of the subparsers are as follows:

Note that the path to save the model is a directory. After training, there are three files in the directory which are named “config.json”, “vocab.pt” and “parser.pt”.

$ python run.py train -h
usage: run.py train [-h] --train_path TRAIN_PATH --dev_path DEV_PATH
                    [--emb_path EMB_PATH] --save_path SAVE_PATH --config_path
                    CONFIG_PATH [--test_wiki_path TEST_WIKI_PATH]
                    [--test_20k_path TEST_20K_PATH] [--gpu GPU] [--seed SEED]
                    [--threads THREADS]
optional arguments:
  -h, --help            show this help message and exit
  --train_path TRAIN_PATH
                        train data dir
  --dev_path DEV_PATH   dev data dir
  --emb_path EMB_PATH   pretrained embedding path
  --save_path SAVE_PATH
                        dic to save all file
  --config_path CONFIG_PATH
                        dic to save all file
  --test_wiki_path TEST_WIKI_PATH
                        wiki test data dir
  --test_20k_path TEST_20K_PATH
                        20k data dir
  --gpu GPU             gpu id
  --seed SEED           random seed
  --threads THREADS     thread num
$ python run.py evaluate -h
usage: run.py evaluate [-h] --gold_path GOLD_PATH --save_path SAVE_PATH
                       [--batch_size BATCH_SIZE] [--gpu GPU] [--seed SEED]
                       [--threads THREADS]
optional arguments:
  -h, --help            show this help message and exit
  --gold_path GOLD_PATH
                        gold test data dir
  --save_path SAVE_PATH
                        path to save the model
  --batch_size BATCH_SIZE
                        batch size
  --gpu GPU             gpu id
  --seed SEED           random seed
  --threads THREADS     thread num
$ python run.py predict -h
usage: run.py predict [-h] --test_path TEST_PATH --save_path SAVE_PATH
                      --pred_path PRED_PATH [--batch_size BATCH_SIZE]
                      [--gpu GPU] [--seed SEED] [--threads THREADS]
optional arguments:
  -h, --help            show this help message and exit
  --test_path TEST_PATH
                        test data dir
  --save_path SAVE_PATH
                        path to save the model
  --pred_path PRED_PATH
                        save predict passages
  --batch_size BATCH_SIZE
                        batch size
  --gpu GPU             gpu id
  --seed SEED           random seed
  --threads THREADS     thread num

Conversion

Conversion code is included in parser.convert. The function UCCA2tree is used to convert a UCCA passage to a tree. The function to_UCCA is used to convert a tree to a UCCA passage. Remote edge recovery code is included in parser.submodel.remote_parser.py independently.