jbrry/g2g-transformer

Pytorch implementation of “Recursive Non-Autoregressive Graph-to-Graph Transformer for Dependency Parsing with Iterative Refinement”

Graph-to-Graph Transformers

Self-attention models, such as Transformer, have been hugely successful in a wide range of natural language processing (NLP) tasks, especially when combined with language-model pre-training, such as BERT.

We propose "Graph-to-Graph Transformer" and "Recursive Non-Autoregressive Graph-to-Graph Transformer for Dependency Parsing with Iterative Refinement"(accepted to TACL) to generalize vanilla Transformer to encode graph structure, and builds the desired output graph.

Note : To use G2GTr model for transition-based dependency parsing, please refer to G2GTr repository.

Contents

• Installation
• Quick Start
• Data Pre-processing and Initial Parser
• Training
• Evaluation
• Predict Raw Sentences
• Citations

Installation

Following packages should be included in your environment:

• Python >= 3.7
• PyTorch >= 1.4.0
• Transformers(huggingface) = 2.4.1

The easier way is to run the following command:

conda env create -f environment.yml
conda activate rngtr

Quick Start

Graph-to-Graph Transformer architecture is general and can be applied to any NLP tasks which interacts with graphs. To use our implementation in your task, you just need to add BertGraphModel class to your code to encode both token-level and graph-level information. Here is a sample usage:

#Loading BertGraphModel and initialize it with available BERT models.
import torch
from parser.utils.graph import initialize_bertgraph,BertGraphModel
# inputing unlabelled graph with label size 5, and Layer Normalization of key
# you can load other BERT pre-trained models too.
encoder = initialize_bertgraph('bert-base-cased',layernorm_key=True,layernorm_value=False,
             input_label_graph=False,input_unlabel_graph=True,label_size=5)

#sample input
input = torch.tensor([[1,2],[3,4]])
graph = torch.tensor([ [[1,0],[0,1]],[[0,1],[1,0]] ])
graph_rel = torch.tensor([[0,1],[3,4]])
output = encoder(input_ids=input,graph_arc=graph,graph_rel=graph_rel)
print(output[0].shape)
## torch.Size([2, 2, 768])

# inputting labelled graph
encoder = initialize_bertgraph('bert-base-cased',layernorm_key=True,layernorm_value=False,
             input_label_graph=True,input_unlabel_graph=False,label_size=5)

#sample input
input = torch.tensor([[1,2],[3,4]])
graph = torch.tensor([ [[2,0],[0,3]],[[0,1],[4,0]] ])
output = encoder(input_ids=input,graph_arc=graph,)
print(output[0].shape)
## torch.Size([2, 2, 768])

If you just want to use BertGraphModel in your research, you can just import it from our repository:

from parser.utils.graph import BertGraphModel,BertGraphConfig
config = BertGraphConfig(YOUR-CONFIG)
config.add_graph_par(GRAPH-CONFIG)
encoder = BertGraphModel(config)

Data Pre-processing and Initial Parser

Dataset Preparation

We evaluated our model on UD Treebanks, English and Chinese Penn Treebanks, and CoNLL 2009 Shared Task. In following sections, we prepare datasets and their evaluation scripts.

Penn Treebanks

English Penn Treebank can be downloaded from english and chinese under LDC license. For English Penn Treebank, replace gold POS tags with Stanford POS tagger with following command in this repository:

bash scripts/postag.sh ${data_dir}/ptb3-wsj-[train|dev|dev.proj|test].conllx

CoNLL 2009 Treebanks

You can download Treebanks from here under LDC license. We use predicted POS tags provided by organizers.

UD Treebanks

You can find required Treebanks from here. (use version 2.3)

Initial Parser

As mentioned in our paper, you can use any initial parser to produce dependency graph. Here we use Biaffine Parser for Penn Treebanks, and German Corpus. We also apply our model to ouput prediction of UDify parser for UD Treebanks.
Biaffine Parser: To prepare biaffine initial parser, we use this repository to produce output predictions.
UDify Parser: For UD Treebanks, we use UDify repository to produce required initial dependency graph.
Alternatively, you can easily run the following command file to produce all required outputs:

bash job_scripts/udify_dataset.bash

Training

To train your own model, you can easily fill out the script in job_scripts directory, and run it. Here is the list of sample scripts:

Model	Script
Syntactic Transformer	baseline.bash
Any initial parser+RNGTr	rngtr.bash
Empty+RNGTr	empty_rngtr.bash

Evaluation

First you should download official scripts from UD, Penn Treebaks, and German. Then, run the following command:

bash job_scripts/predict.bash

To replicate refinement analysis and error analysis results, you should use MaltEval tools.

Predict Raw Sentences

You can also predict dependency graphs of raw texts with a pre-trained model by modifying predict.bash file. Just set input_type to raw. Then, put all your sentences in a .txt file, and the output will be in CoNNL format.

Citations

If you use this code for your research, please cite these works as:

@misc{mohammadshahi2020recursive,
      title={Recursive Non-Autoregressive Graph-to-Graph Transformer for Dependency Parsing with Iterative Refinement}, 
      author={Alireza Mohammadshahi and James Henderson},
      year={2020},
      eprint={2003.13118},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@inproceedings{mohammadshahi-henderson-2020-graph,
    title = "Graph-to-Graph Transformer for Transition-based Dependency Parsing",
    author = "Mohammadshahi, Alireza  and
      Henderson, James",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.294",
    pages = "3278--3289",
    abstract = "We propose the Graph2Graph Transformer architecture for conditioning on and predicting arbitrary graphs, and apply it to the challenging task of transition-based dependency parsing. After proposing two novel Transformer models of transition-based dependency parsing as strong baselines, we show that adding the proposed mechanisms for conditioning on and predicting graphs of Graph2Graph Transformer results in significant improvements, both with and without BERT pre-training. The novel baselines and their integration with Graph2Graph Transformer significantly outperform the state-of-the-art in traditional transition-based dependency parsing on both English Penn Treebank, and 13 languages of Universal Dependencies Treebanks. Graph2Graph Transformer can be integrated with many previous structured prediction methods, making it easy to apply to a wide range of NLP tasks.",
}

Have a question not listed here? Open a GitHub Issue or send us an email.