/graph4nlp

Graph4nlp is the library for the easy use of Graph Neural Networks for NLP. Welcome to visit our DLG4NLP website (https://dlg4nlp.github.io/index.html) for various learning resources!

Primary LanguagePythonApache License 2.0Apache-2.0

logo

Last Commit pypi Contributors Contributing License Issues Fork Star

Graph4NLP

Graph4NLP is an easy-to-use library for R&D at the intersection of Deep Learning on Graphs and Natural Language Processing (i.e., DLG4NLP). It provides both full implementations of state-of-the-art models for data scientists and also flexible interfaces to build customized models for researchers and developers with whole-pipeline support. Built upon highly-optimized runtime libraries including DGL , Graph4NLP has both high running efficiency and great extensibility. The architecture of Graph4NLP is shown in the following figure, where boxes with dashed lines represents the features under development. Graph4NLP consists of four different layers: 1) Data Layer, 2) Module Layer, 3) Model Layer, and 4) Application Layer.

architecture
Figure: Graph4NLP Overall Architecture

new Graph4NLP news

01/20/2022: The v0.5.5 release. Try it out!
09/26/2021: The v0.5.1 release. Try it out!
09/01/2021: Welcome to visit our DLG4NLP website (https://dlg4nlp.github.io/index.html) for various learning resources!
06/05/2021: The v0.4.1 release.

Major Releases

Releases Date Features
v0.5.5 2022-01-20 - Support model.predict API by introducing wrapper functions.
- Introduce Three new inference_wrapper functions: classifier_inference_wrapper, generator_inference_wrapper, generator_inference_wrapper_for_tree.
- Add the inference and inference_advance examples in each application.
- Separate the graph topology and graph embedding process.
- Renew all the graph construction functions.
- Module graph_embedding is divided into graph_embedding_initialization and graph_embedding_learning.
- Unify the parameters in Dataset. We remove the ambiguous parameter graph_type and introduce graph_name to indicate the graph construction method and static_or_dynamic to indicate the static or dynamic graph construction type.
- New: The dataset now can automatically choose the default methods (e.g., topology_builder) by only one parameter graph_name.
v0.5.1 2021-09-26 - Lint the codes
- Support testing with users' own data
- Fix the bug: The word embedding size was hard-coded in the 0.4.1 version. Now it is equal to "word_emb_size" parameter.
- Fix the bug: The build_vocab() is called twice in the 0.4.1 version.
- Fix the bug: The two main files of knowledge graph completion example missed the optional parameter "kg_graph" in ranking_and_hits() when resuming training the model.
- Fix the bug: We have fixed the preprocessing path error in KGC readme.
- Fix the bug: We have fixed embedding construction bug when setting emb_strategy to 'w2v'.
v0.4.1 2021-06-05 - Support the whole pipeline of Graph4NLP
- GraphData and Dataset support

Quick tour

Graph4nlp aims to make it incredibly easy to use GNNs in NLP tasks (check out Graph4NLP Documentation). Here is an example of how to use the Graph2seq model (widely used in machine translation, question answering, semantic parsing, and various other NLP tasks that can be abstracted as graph-to-sequence problem and has shown superior performance).

We also offer other high-level model APIs such as graph-to-tree models. If you are interested in DLG4NLP related research problems, you are very welcome to use our library and refer to our graph4nlp survey.

from graph4nlp.pytorch.datasets.jobs import JobsDataset
from graph4nlp.pytorch.modules.graph_construction.dependency_graph_construction import DependencyBasedGraphConstruction
from graph4nlp.pytorch.modules.config import get_basic_args
from graph4nlp.pytorch.models.graph2seq import Graph2Seq
from graph4nlp.pytorch.modules.utils.config_utils import update_values, get_yaml_config

# build dataset
jobs_dataset = JobsDataset(root_dir='graph4nlp/pytorch/test/dataset/jobs',
                           topology_builder=DependencyBasedGraphConstruction,
                           topology_subdir='DependencyGraph')  # You should run stanfordcorenlp at background
vocab_model = jobs_dataset.vocab_model

# build model
user_args = get_yaml_config("examples/pytorch/semantic_parsing/graph2seq/config/dependency_gcn_bi_sep_demo.yaml")
args = get_basic_args(graph_construction_name="node_emb", graph_embedding_name="gat", decoder_name="stdrnn")
update_values(to_args=args, from_args_list=[user_args])
graph2seq = Graph2Seq.from_args(args, vocab_model)

# calculation
batch_data = JobsDataset.collate_fn(jobs_dataset.train[0:12])

scores = graph2seq(batch_data["graph_data"], batch_data["tgt_seq"])  # [Batch_size, seq_len, Vocab_size]

Overview

Our Graph4NLP computing flow is shown as below.

logo

Graph4NLP Models and Applications

Graph4NLP models

  • Graph2Seq: a general end-to-end neural encoder-decoder model that maps an input graph to a sequence of tokens.
  • Graph2Tree: a general end-to-end neural encoder-decoder model that maps an input graph to a tree structure.

Graph4NLP applications

We provide a comprehensive collection of NLP applications, together with detailed examples as follows:

  • Text classification: to give the sentence or document an appropriate label.
  • Semantic parsing: to translate natural language into a machine-interpretable formal meaning representation.
  • Neural machine translation: to translate a sentence in a source language to a different target language.
  • summarization: to generate a shorter version of input texts which could preserve major meaning.
  • KG completion: to predict missing relations between two existing entities in konwledge graphs.
  • Math word problem solving: to automatically solve mathematical exercises that provide background information about a problem in easy-to-understand language.
  • Name entity recognition: to tag entities in input texts with their corresponding type.
  • Question generation: to generate an valid and fluent question based on the given passage and target answer (optional).

Performance

Environment: torch 1.8, ubuntu 16.04 with 2080ti GPUs

Task Dataset GNN Model Graph construction Evaluation Performance
Text classification TRECT
CAirline
CNSST
GAT Dependency
Constituency
Dependency
Accuracy 0.948
0.785
0.538
Semantic Parsing JOBS SAGE Constituency Execution accuracy 0.936
Question generation SQuAD GGNN Dependency BLEU-4 0.15175
Machine translation IWSLT14 GCN Dynamic BLEU-4 0.3212
Summarization CNN(30k) GCN Dependency ROUGE-1 26.4
Knowledge graph completion Kinship GCN Dependency MRR 82.4
Math word problem MAWPS SAGE Dynamic Solution accuracy 76.4

Installation

Currently, users can install Graph4NLP via pip or source code. Graph4NLP supports the following OSes:

  • Linux-based systems (tested on Ubuntu 18.04 and later)
  • macOS (only CPU version)
  • Windows 10 (only support pytorch >= 1.8)

Installation via pip (binaries)

We provide pip wheels for all major OS/PyTorch/CUDA combinations. Note that we highly recommend Windows users refer to Installation via source code due to compatibility.

Ensure that at least PyTorch (>=1.6.0) is installed:

Note that >=1.6.0 is ok.

$ python -c "import torch; print(torch.__version__)"
>>> 1.6.0

Find the CUDA version PyTorch was installed with (for GPU users):

$ python -c "import torch; print(torch.version.cuda)"
>>> 10.2

Install the relevant dependencies:

torchtext is needed since Graph4NLP relies on it to implement embeddings. Please pay attention to the PyTorch requirements before installing torchtext with the following script! For detailed version matching please refer here.

pip install torchtext # >=0.7.0

Install Graph4NLP

pip install graph4nlp${CUDA}

where ${CUDA} should be replaced by the specific CUDA version (none (CPU version), "-cu92", "-cu101", "-cu102", "-cu110"). The following table shows the concrete command lines. For CUDA 11.1 users, please refer to Installation via source code.

Platform Command
CPU pip install graph4nlp
CUDA 9.2 pip install graph4nlp-cu92
CUDA 10.1 pip install graph4nlp-cu101
CUDA 10.2 pip install graph4nlp-cu102
CUDA 11.0 pip install graph4nlp-cu110

Installation via source code

Ensure that at least PyTorch (>=1.6.0) is installed:

Note that >=1.6.0 is ok.

$ python -c "import torch; print(torch.__version__)"
>>> 1.6.0

Find the CUDA version PyTorch was installed with (for GPU users):

$ python -c "import torch; print(torch.version.cuda)"
>>> 10.2

Install the relevant dependencies:

torchtext is needed since Graph4NLP relies on it to implement embeddings. Please pay attention to the PyTorch requirements before installing torchtext with the following script! For detailed version matching please refer here.

pip install torchtext # >=0.7.0

Download the source code of Graph4NLP from Github:

git clone https://github.com/graph4ai/graph4nlp.git
cd graph4nlp

Configure the CUDA version

Then run ./configure (or ./configure.bat if you are using Windows 10) to config your installation. The configuration program will ask you to specify your CUDA version. If you do not have a GPU, please type 'cpu'.

./configure

Install the relevant packages:

Finally, install the package:

python setup.py install

For Hyperparameter tuning

We show some of the hyperparameters that are often tuned here.

New to Deep Learning on Graphs for NLP?

If you want to learn more on applying Deep Learning on Graphs techniques to NLP tasks, welcome to visit our DLG4NLP website (https://dlg4nlp.github.io/index.html) for various learning resources! You can refer to our survey paper which provides an overview of this existing research direction. If you want detailed reference to our library, please refer to our docs.

Contributing

Please let us know if you encounter a bug or have any suggestions by filing an issue.

We welcome all contributions from bug fixes to new features and extensions.

We expect all contributions discussed in the issue tracker and going through PRs.

Citation

If you found this code useful, please consider citing the following papers.

@article{wu2021graph,
  title={Graph Neural Networks for Natural Language Processing: A Survey},
  author={Lingfei Wu and Yu Chen and Kai Shen and Xiaojie Guo and Hanning Gao and Shucheng Li and Jian Pei and Bo Long},
  journal={arXiv preprint arXiv:2106.06090},
  year={2021}
}

@inproceedings{chen2020iterative,
  title={Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings},
  author={Chen, Yu and Wu, Lingfei and Zaki, Mohammed J},
  booktitle={Proceedings of the 34th Conference on Neural Information Processing Systems},
  month={Dec. 6-12,},
  year={2020}
}

@inproceedings{chen2020reinforcement,
  author    = {Chen, Yu and Wu, Lingfei and Zaki, Mohammed J.},
  title     = {Reinforcement Learning Based Graph-to-Sequence Model for Natural Question Generation},
  booktitle = {Proceedings of the 8th International Conference on Learning Representations},
  month = {Apr. 26-30,},
  year      = {2020}
}

@article{xu2018graph2seq,
  title={Graph2seq: Graph to sequence learning with attention-based neural networks},
  author={Xu, Kun and Wu, Lingfei and Wang, Zhiguo and Feng, Yansong and Witbrock, Michael and Sheinin, Vadim},
  journal={arXiv preprint arXiv:1804.00823},
  year={2018}
}

@inproceedings{li-etal-2020-graph-tree,
    title = {Graph-to-Tree Neural Networks for Learning Structured Input-Output Translation with Applications to Semantic Parsing and Math Word Problem},
    author = {Li, Shucheng  and
      Wu, Lingfei  and
      Feng, Shiwei  and
      Xu, Fangli  and
      Xu, Fengyuan  and
      Zhong, Sheng},
    booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2020},
    month = {Nov},
    year = {2020}
}

@inproceedings{huang-etal-2020-knowledge,
    title = {Knowledge Graph-Augmented Abstractive Summarization with Semantic-Driven Cloze Reward},
    author = {Huang, Luyang  and
      Wu, Lingfei  and
      Wang, Lu},
    booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
    month = {Jul},
    year = {2020},
    pages = {5094--5107}
}

@inproceedings{wu-etal-2018-word,
    title = {Word Mover{'}s Embedding: From {W}ord2{V}ec to Document Embedding},
    author = {Wu, Lingfei  and
      Yen, Ian En-Hsu  and
      Xu, Kun  and
      Xu, Fangli  and
      Balakrishnan, Avinash  and
      Chen, Pin-Yu  and
      Ravikumar, Pradeep  and
      Witbrock, Michael J.},
    booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
    pages = {4524--4534},
    year = {2018},
}

@inproceedings{chen2020graphflow,
  author    = {Yu Chen and
               Lingfei Wu and
               Mohammed J. Zaki},  
title     = {GraphFlow: Exploiting Conversation Flow with Graph Neural Networks
               for Conversational Machine Comprehension},
  booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI} 2020},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  pages     = {1230--1236},
  year      = {2020}
} 
  
@inproceedings{shen2020hierarchical,
  title={Hierarchical Attention Based Spatial-Temporal Graph-to-Sequence Learning for Grounded Video Description},
  author={Shen, Kai and Wu, Lingfei and Xu, Fangli and Tang, Siliang and Xiao, Jun and Zhuang, Yueting},
  booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI} 2020},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  pages     = {941--947},
  year      = {2020}
}  

@inproceedings{ijcai2020-419,
  title     = {RDF-to-Text Generation with Graph-augmented Structural Neural Encoders},
  author    = {Gao, Hanning and Wu, Lingfei and Hu, Po and Xu, Fangli},
  booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI-20}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  pages     = {3030--3036},
  year      = {2020}
}


Team

Graph4AI Team: Lingfei Wu (team leader), Yu Chen, Kai Shen, Xiaojie Guo, Hanning Gao, Shucheng Li, Saizhuo Wang, Xiao Liu and Jing Hu. We are passionate in developing useful open-source libraries which aim to promote the easy use of various Deep Learning on Graphs techniques for Natural Language Processing. Our team consists of research scientists, applied data scientists, and graduate students from a variety of industrial and academic groups, including Pinterest (Lingfei Wu), Zhejiang University (Kai Shen), Facebook AI (Yu Chen), IBM T.J. Watson Research Center (Xiaojie Guo), Tongji University (Hanning Gao), Nanjing University (Shucheng Li), HKUST (Saizhuo Wang).

Contact

If you have any technical questions, please submit new issues.

If you have any other questions, please contact us: Lingfei Wu [lwu@email.wm.edu] and Xiaojie Guo [xiaojie.guo@jd.com].

License

Graph4NLP uses Apache License 2.0.