Transition-based UCCA Parser

TUPA is a transition-based parser for Universal Conceptual Cognitive Annotation (UCCA).

Requirements

Python 3.6+

Install

Install the latest release:

pip install tupa[bert]

Alternatively, install the latest code from GitHub (may be unstable):

pip install git+git://github.com/danielhers/tupa.git#egg=tupa

Train the parser

Having a directory with UCCA passage files (for example, the English Wiki corpus), run:

python -m tupa -t <train_dir> -d <dev_dir> -c <model_type> -m <model_filename>

The possible model types are sparse, mlp, and bilstm.

Parse a text file

Run the parser on a text file (here named example.txt) using a trained model:

python -m tupa example.txt -m <model_filename>

An xml file will be created per passage (separate by blank lines in the text file).

To visualize output graphs, use the visualize command:

python -m scripts.visualize <xml_files>

Pre-trained models

To download and extract a model pre-trained on the Wiki corpus, run:

curl -LO https://github.com/huji-nlp/tupa/releases/download/v1.3.10/ucca-bilstm-1.3.10.tar.gz
tar xvzf ucca-bilstm-1.3.10.tar.gz

Run the parser using the model:

python -m tupa example.txt -m models/ucca-bilstm

Other languages

To get a model pre-trained on the French 20K Leagues corpus or a model pre-trained on the German 20K Leagues corpus, run:

curl -LO https://github.com/huji-nlp/tupa/releases/download/v1.3.10/ucca-bilstm-1.3.10-fr.tar.gz
tar xvzf ucca-bilstm-1.3.10-fr.tar.gz
curl -LO https://github.com/huji-nlp/tupa/releases/download/v1.3.10/ucca-bilstm-1.3.10-de.tar.gz
tar xvzf ucca-bilstm-1.3.10-de.tar.gz

Run the parser on a French/German text file (separate passages by blank lines):

python -m tupa exemple.txt -m models/ucca-bilstm-fr --lang fr
python -m tupa beispiel.txt -m models/ucca-bilstm-de --lang de

Using BERT

BERT can be used instead of standard word embeddings. First, install the required dependencies:

pip install tupa[bert]

or, if you cloned the repository,

pip install -r requirements.bert.txt

Then pass the --use-bert argument to the training command.

See the possible configuration options in config.py (relevant options have the prefix bert).

BERT Multilingual Training

A multilingual model can be trained, to leverage cross-lingual transfer and improve results on low-resource languages:

Make sure the input passage files have the lang attribute. See the script set_lang in the package semstr.
Enable BERT by passing the --use-bert argument.
Use the multilingual model by passing --bert-model=bert-base-multilingual-cased.
Pass the --bert-multilingual=0 argument to enable multilingual training.

BERT Performance

Here are the average results over 3 BERT multilingual models trained on the German 20K Leagues corpus, English Wiki corpus and only on 15 sentences from the French 20K Leagues corpus, with the following settings:

bert-model=bert-base-multilingual-cased
bert-layers=-1 -2 -3 -4
bert-layers-pooling=weighted
bert-token-align-by=sum

The results:

description	test primary F1	test remote F1	test average
German 20K Leagues	0.828	0.6723	0.824
English 20K Leagues	0.763	0.359	0.755
French 20K Leagues	0.739	0.46	0.732
English Wiki	0.789	0.581	0.784

*English 20K Leagues corpus is used as out of domain test.

Pre-trained Models with BERT

To download and extract a multilingual model trained with the settings above, run:

curl -LO https://github.com/huji-nlp/tupa/releases/download/v1.4.0/bert_multilingual_layers_4_layers_pooling_weighted_align_sum.tar.gz
tar xvzf bert_multilingual_layers_4_layers_pooling_weighted_align_sum.tar.gz

To run the parser using the model, use the following command. Pay attention that you need to replace [lang] with the right language symbol (fr, en, or de):

python -m tupa example.txt --lang [lang] -m bert_multilingual_layers_4_layers_pooling_weighted_align_sum

Author

Daniel Hershcovich: dh@di.ku.dk

Contributors

Ofir Arviv: ofir.arviv@mail.huji.ac.il

Citation

If you make use of this software, please cite the following paper:

@InProceedings{hershcovich2017a,
  author    = {Hershcovich, Daniel  and  Abend, Omri  and  Rappoport, Ari},
  title     = {A Transition-Based Directed Acyclic Graph Parser for {UCCA}},
  booktitle = {Proc. of ACL},
  year      = {2017},
  pages     = {1127--1138},
  url       = {http://aclweb.org/anthology/P17-1104}
}

The version of the parser used in the paper is v1.0. To reproduce the experiments, run:

curl -L https://raw.githubusercontent.com/huji-nlp/tupa/master/experiments/acl2017.sh | bash

If you use the French, German or multitask models, please cite the following paper:

@InProceedings{hershcovich2018multitask,
  author    = {Hershcovich, Daniel  and  Abend, Omri  and  Rappoport, Ari},
  title     = {Multitask Parsing Across Semantic Representations},
  booktitle = {Proc. of ACL},
  year      = {2018},
  pages     = {373--385},
  url       = {http://aclweb.org/anthology/P18-1035}
}

The version of the parser used in the paper is v1.3.3. To reproduce the experiments, run:

curl -L https://raw.githubusercontent.com/huji-nlp/tupa/master/experiments/acl2018.sh | bash

License

This package is licensed under the GPLv3 or later license (see LICENSE.txt).

            [ ~ Dependencies scanned by PyUp.io ~ ]