/AMR-IE

The code repository for AMR guided joint information extraction model (NAACL-2021).

Primary LanguagePython

AMR-IE

The code repository for AMR guided joint information extraction model (NAACL-HLT 2021).

AMR Parser

We use the transformer-based AMR parser (Astudillo et al. 2020) in our model. We train the AMR parser on LDC AMR 3.0 data, and the code and the pretrained model are publicly available. Please follow the README file in the code directory to install and use it.

The code of the AMR parser is derived from the original Github Repo. We made some slight modifications to fit it in higher versions of PyTorch.

Environments

The core python dependencies for this code are as follows:

torch==1.6.0, torch-scatter==2.0.6, transformers==4.3.0, dgl==0.6.0.

For other dependencies, you can simply use pip install to install the required packages that are indicated by the error messages.

And for IBM/transition-amr-parser's dependencies, refer to this link, DO NOT pip install directly.

关于使用IBM的pretrained-parser（stack-transformer）有一些需要记录的点：

  • 目前已经可以使用，并跑通了对oneie数据的parse。
  • 原repo是没有也不会放出来pretrained-model的，因为issue里面他们说训练用的AMR LDC数据集是非公开的，所以不放。
  • 但是AMR-IE repo放出来了pretrained-parser，包括所使用的code和pretrained-model(checkpoint)，pretrained-model是在AMR-IE作者在AMR3.0上训练好的（We used IBM’s code and retrained the AMR parser using AMR 3.0 data (LDC2019E81).）。使用的时候要follow原repo给出的installation instructions才能装好，跑的时候要load checkpoint。
  • 想要3090跑起来这个parser，依赖怎么安：首先按照AMR-IE的README安装，torch因为要3090，所以去用兼容cuda11以上的版本，他要求1.6不兼容3090，我用1.10也没报错，不过版本不同终究可能会导致复现结果的不同吧。其余三个包用pip安就行，然后安一下lxml和nltk。最重要的是parser相关依赖的安装，不能直接去pip或者conda fairseq，而是要follow原repo给出的installation instructions从本地对fairseq进行patch后再安装，具体follow原repo也就够了。
  • 关于生成的数据：我用两台机器都对他们给出的genia2011的oneiejson文件进行parse，跑出来的结果是同样的，这证明用pretrained parser进行amr parse没有随机性，但是这两个结果都与他们放出来的预处理好的文件不一，我想有可能是因为依赖版本不同？这个应该影响不大。

Datasets

ACE-2005 and ERE

The ACE-2005 and ERE datasets are only available at the LDC website. Please use the following steps for preprocessing and AMR parsing.

  • Use the scripts in OneIE to obtain the data files train.oneie.json, dev.oneie.json, and test.oneie.json for the OneIE model.
  • Use data/transform_for_amrie.py to transform the OneIE formatted data files to fit in our model. For example: python ./data/transform_for_amrie.py -i [INPUT_DATA_DIR] -o [OUTPUT_DATA_DIR].
  • Use ./process_amr.py to conduct AMR parsing and generate the .pkl files for training the model.

After preprocessing, each dataset split should contain a .json file (containing the IE annotations) and a .pkl file (containing AMR information). For example, for the training set of ACE-2005, we have train.oneie.json and train_graphs.pkl. You can also refer to the publicly available GENIA datasets for more detailed dataset formats in our model.

GENIA 2011 and 2013

We release the preprocessed data for GENIA 2011 and 2013 along with the AMR graphs for each sentence at this link. Please unzip the file and put all the folders into ./data/ before training the models.

Training

To train a model, you can run python train.py -c [CONFIG_FILE] -g [GPU_ID] -n [NAME_OF_THIS_RUN]. The example config files are provided in ./config/. For example, to train a model on GENIA 2011 joint IE dataset, you can run:

python train.py -c config/genia_2011.json -g 0 -n your_preferred_name

The average training time for a single model would be ~15 hours on Tesla V100 GPUs, with ~10GB usage of GPU memory.

Acknowledgement

This code is largely derived from OneIE. Our great thanks to Lin et al. and Astudillo et al. for publicizing their codes for the OneIE model and the AMR parser!

Please contact zixuan11@illinois.edu if you have any questions. If you use this code as part of your research, please cite the following paper:

@inproceedings{amrie2021,
  author    = {Zixuan Zhang and Heng Ji},
  title     = {Abstract Meaning Representation Guided Graph Encoding and Decodingfor Joint Information Extraction},
  booktitle = {Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT)},
  year      = {2021},
  pages     = {39--49},
  url       = {https://www.aclweb.org/anthology/2021.naacl-main.4/}
  }