Skip to content

Latest commit

 

History

History
228 lines (199 loc) · 8.23 KB

README.md

File metadata and controls

228 lines (199 loc) · 8.23 KB
Raw

CRFIE

Python ArXiv ACL

This repo contains the code used for the ACL 2023 paper Modeling Instance Interactions for Joint Information Extraction with Neural High-Order Conditional Random Field.

Requirements

  • Python 3.7
  • Python packages
    • PyTorch 1.0+ (Install the CPU version if you use this tool on a machine without GPUs)
    • transformers 3.0.2 (Using transformers 3.1+ may cause some model loading issue)
    • tqdm
    • lxml
    • nltk

Getting Started

Pre-processing

Our preprocessing mainly adapts from OneIE.

Preprocess DyGIE++

The prepreocessing/process_dygiepp.py script converts datasets in DyGIE+ format to the input format.
Example:

python preprocessing/process_dygiepp.py -i train.json -o train.oneie.json

Arguments:

  • -i, --input: Path to the input file.
  • -o, --output: Path to the output file.

Preprocess ACE2005

The prepreocessing/process_ace.py script converts raw ACE2005 datasets to the input format. Example:

python preprocessing/process_ace.py -i <INPUT_DIR>/LDC2006T06/data -o <OUTPUT_DIR> \
      -s resource/splits/ACE05-E -b bert-large-cased -c <BERT_CACHE_DIR> -l english

Arguments:

  • -i, --input: Path to the input directory (data folder in your LDC2006T06 package).
  • -o, --output: Path to the output directory.
  • -b, --bert: Bert model name.
  • -c, --bert_cache_dir: Path to the BERT cache directory.
  • -s, --split: Path to the split directory. We provide document id lists for all datasets used in our paper in resource/splits.
  • -l, --lang: Language (options: english, chinese).

Preprocess ERE

The prepreocessing/process_ere.py script converts raw ERE datasets (LDC2015E29, LDC2015E68, LDC2015E78, LDC2015E107) to the input format.

python preprocessing/process_ere.py -i <INPUT_DIR>/data -o <OUTPUT_DIR> -b bert-large-cased -c <BERT_CACHE_DIR> -l english -d normal

Arguments:

  • -i, --input: Path to the input directory (data folder in your ERE package).
  • -o, --output: Path to the output directory.
  • -b, --bert: Bert model name.
  • -c, --bert_cache_dir: Path to the BERT cache directory.
  • -d, --dataset: Dataset type: normal, r2v2, parallel, or spanish.
  • -l, --lang: Language (options: english, spanish).

This script currently supports:

  • LDC2015E29_DEFT_Rich_ERE_English_Training_Annotation_V1
  • LDC2015E29_DEFT_Rich_ERE_English_Training_Annotation_V2
  • LDC2015E68_DEFT_Rich_ERE_English_Training_Annotation_R2_V2
  • LDC2015E78_DEFT_Rich_ERE_Chinese_and_English_Parallel_Annotation_V2
  • LDC2015E107_DEFT_Rich_ERE_Spanish_Annotation_V2

Training

  • cd to the root directory of this package
  • Set the environment variable PYTHONPATH to the current directory. For example, if you unpack this package to ~/High-order-IE, run: export PYTHONPATH=~/High-order-IE

Because our framework is a pipeline schema, you should first train the Node Identification model and save the checkpoint in a directory.

  • Run the command line to train an identification model: python train_ident.py -c <CONFIG_FILE_PATH>.

Then train the high-order classification model.

  • python train.py -c <CONFIG_FILE_PATH>.
  • One example configuration file is in config/baseline.json. Fill in the following paths in the configuration file:
    • BERT_CACHE_DIR: Pre-trained BERT models, configs, and tokenizers will be downloaded to this directory.
    • TRAIN_FILE_PATH, DEV_FILE_PATH, TEST_FILE_PATH: Path to the training/dev/test/files.
    • OUTPUT_DIR: The model will be saved to subfolders in this directory.
    • VALID_PATTERN_DIR: Valid patterns created based on the annotation guidelines or training set. Example files are provided in resource/valid_patterns.
    • Set NER_SCORE and SPLIT_TRAIN to be true: Our base pipeline model with different scoring functions of OneIE.
    • IDENT_MODEL_PATH: Path to a checkpoint of the saved node identification model. The following hyper-parameters control the high-order part:
    • USE_*: Whether to use the corresponding high-order factor.
    • DECOMP_SIZE and MFVI_ITER: Hyperparameters of mean field variational inference (refer to paper).

Evaluation

Example command line to test a file input: python predict.py -m <best.role.mdl> -i <input_dir> -o <output_dir> --format json

  • Arguments:
    • -m, --model_path: Path to the trained model.
    • -i, --input_dir: Path to the input directory. LTF format sample files can be found in the input directory.
    • -o, --output_dir: Path to the output directory (json format). Output files are in the JSON format. Sample files can be found in the output directory.
    • --gpu: (optional) Use GPU
    • -d, --device: (optional) GPU device index (for multi-GPU machines).
    • -b, --batch_size: (optional) Batch size. For a 16GB GPU, a batch size of 10~15 is a reasonable value.
    • --max_len: (optional) Max sentence length. Sentences longer than this value will be ignored. You may need to decrease batch_size if you set max_len to a larger number.
    • --lang: (optional) Model language.
    • --format: Input file format (txt, ltf, or json).

Data Format

Processed input example:

{
    "doc_id": "AFP_ENG_20030401.0476",
    "sent_id": "AFP_ENG_20030401.0476-5",
    "entity_mentions": [
        {
            "id": "AFP_ENG_20030401.0476-5-E0",
            "start": 0,
            "end": 1,
            "entity_type": "GPE",
            "mention_type": "UNK",
            "text": "British"
        },
        ...
    ],
    "relation_mentions": [
        {
            "relation_type": "ORG-AFF",
            "id": "AFP_ENG_20030401.0476-5-R0",
            "arguments": [
                {
                    "entity_id": "AFP_ENG_20030401.0476-5-E1",
                    "text": "Chancellor",
                    "role": "Arg-1"
                },
                ...
            ]
        },
        ...
    ],
    "event_mentions": [
        {
            "event_type": "Personnel:Nominate",
            "id": "AFP_ENG_20030401.0476-5-EV0",
            "trigger": {
                "start": 9,
                "end": 10,
                "text": "named"
            },
            "arguments": [
                {
                    "entity_id": "AFP_ENG_20030401.0476-5-E4",
                    "text": "head",
                    "role": "Person"
                }
            ]
        }
    ],
    "tokens": [
        ...
    ],
    "pieces": [
        ...
    ],
    "token_lens": [
        ...
    ],
    "sentence": ...
}

The "start" and "end" of entities and triggers are token indices. The "arguments" of a relation refer to its head entity and tail entity.

Output example:

{
    "doc_id": "HC0003PYD",
    "sent_id": "HC0003PYD-16",
    "token_ids": [
        ...
    ],
    "tokens": [
        ...
    ],
    "graph": {
        "entities": [
            [
                3,
                5,
                "GPE",
                "NAM",
                1.0
            ],
            ...
        ],
        "triggers": [
            ...
        ],
        "relations": [
            ...
        ],
        "roles": [
            ...
        ]
    }
}

Citation

@inproceedings{jia-etal-2023-modeling,
    title = "Modeling Instance Interactions for Joint Information Extraction with Neural High-Order Conditional Random Field",
    author = "Jia, Zixia and Yan, Zhaohui and Han, Wenjuan and Zheng, Zilong and Tu, Kewei",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.766",
    doi = "10.18653/v1/2023.acl-long.766",
    pages = "13695--13710"
}

Acknowledgments

The codebase of this repo is extended from OneIE v0.4.8