Second Order Parser for Semantic Dependency Parsing

This repo contains the code forked from Parser-v3 and used for the semantic dependency parser in Wang et al. (2019), Second-Order Semantic Dependency Parsing with End-to-End Neural Networks and CoNLL 2019 shared task (SDP part only).

The PyTorch Version for the Second-Order SDP parser is now available at MultilangStructureKD!

python3
tensorflow-gpu>=1.12.0

Our second order parser can be trained by simply running

python3 -u main.py train GraphParserNetwork --config_file config/sec_order.cfg --noscreen

This config file will run Mean Field Variational Inference for second order parts, and if you want to run with Loopy Belief Propagation, run

python3 -u main.py train GraphParserNetwork --config_file config/sec_order_LBP.cfg --noscreen

Our second order parser can be trained with bert. First clone the bert repository:

git clone https://github.com/google-research/bert

Download bert model: BERT-Large, Uncased (Whole Word Masking)

To train with bert, simply run

python3 -u main.py train GraphParserNetwork --config_file config_gen/bert_large_glove_previous_layer_100linear_01lr_5decay_dm_switch_new1.cfg --noscreen

If you want to fine tune bert model, set is_training=True in BertVocab

A trained model can be run by calling

python3 main.py --save_dir $SAVEDIR run $DATADIR --output_dir results 

The parsed result will be saved results/ directory. The $SAVEDIR is the directory of the model, for example, if you trained with config/sec_order.cfg, the model will be saved in saves/SemEval15/DM/MF_dm_3iter. The $DATADIR is the directory of the data in CONLLU format.

The pretrained model on DM can be download from the following links:

Baidu Netdisk, Password: ecqe

Google Drive

The model is trained with Bert and Glove embeddings considering there are no golden POS tags and lemmas in practice, the Labeled F1 score is 94.25 and 90.76 for in-domain and out-of-domain respectively.

To avoid out of memory in training phase, our parser should be trained with 12GB gpu memory, and no longer than 60 words for each sentence. The number of iterations for mean field variational inference is at most 3 and at most 2 for loopy belief propagation in a 12GB Titan X gpu. If you have a larger gpu, such as Tesla P40 24GB, loopy belief propation can be also trained with 3 iterations. To set the number of iterations, set num_iteration in SecondOrderGraphIndexVocab or SecondOrderGraphLBPVocab of the config file. Another way is reduce the training batch_size in CoNLLUTrainset of the config file.

If you want to see some details of our parser, the source code for our parser is in parser/structs/vocabs/second_order_vocab.py for Mean Field Variational Inference and second_order_LBP_vocab.py for Loopy Belief Propagation in the same directory.

If you find our code is useful, please cite:

@inproceedings{wang-etal-2019-second,
    title = "Second-Order Semantic Dependency Parsing with End-to-End Neural Networks",
    author = "Wang, Xinyu  and
      Huang, Jingxian  and
      Tu, Kewei",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1454",
    pages = "4609--4618",}

@inproceedings{Wan:Liu:Jia:19,
  author = {Wang, Xinyu and Liu, Yixian and Jia, Zixia
            and Jiang, Chengyue and Tu, Kewei},
  title = {{ShanghaiTech} at {MRP}~2019:
           {S}equence-to-Graph Transduction with Second-Order Edge Inference
           for Cross-Framework Meaning Representation Parsing},
  booktitle = CONLL:19:U,
  address = L:CONLL:19,
  pages = {\pages{--}{55}{65}},
  year = 2019
}

If you have any questions, feel free to contact with me through email.