AtTGen

Implementation of "AtTGen: Attribute Tree Generation for Real-World Attribute Joint Extraction", ACL 2023.

A lightweight attribute extraction model that achieves above 96% F1-score on MEPAVE dataset ;)

Preparation

Please install the dependencies first:

pip install -r requirements.txt

Usage

usage: main.py [-h] [--name NAME] [--do_train] [--do_eval]
               [--data_dir DATA_DIR] [--word_vocab WORD_VOCAB]
               [--ontology_vocab ONTOLOGY_VOCAB] [--tokenizer TOKENIZER]
               [--seed SEED] [--gpu_ids GPU_IDS] [--batch_size BATCH_SIZE]
               [--lr LR] [--epoch EPOCH] [--emb_dim EMB_DIM]
               [--encode_dim ENCODE_DIM] [--skip_subject SKIP_SUBJECT]

configuration

optional arguments:
  -h, --help            show this help message and exit
  --name NAME           Experiment name, for logging and saving models
  --do_train            Whether to run training.
  --do_eval             Whether to run eval on the test set.
  --data_dir DATA_DIR   The input data dir.
  --word_vocab WORD_VOCAB
                        The vocabulary file.
  --ontology_vocab ONTOLOGY_VOCAB
                        The ontology class file.
  --tokenizer TOKENIZER
                        The tokenizer type.
  --seed SEED           The random seed for initialization
  --gpu_ids GPU_IDS     The GPU ids
  --batch_size BATCH_SIZE
                        Total batch size for training.
  --lr LR               The initial learning rate for Adam.
  --epoch EPOCH         Total number of training epochs to perform.
  --emb_dim EMB_DIM     The dimension of the embedding
  --encode_dim ENCODE_DIM
                        The dimension of the encoding
  --skip_subject SKIP_SUBJECT
                        Whether to skip the subject

Dataset

Download the dataset to the raw_data folder, and run python3 preprocess.py --dataset=xxxx to preprocess the data.

Using argument --subject_guild True to enable the subject guild function.

Pre-processed NYT dataset is attached in the data folder, which can be used directly.

Playground (Recommended)

Benefiting from the parameter-efficiency of this model, we can easily train and inference the model, and evaluate the trained model weights conveniently.

The trained model weights are in runs/jave_best file, which is trained by default hyper-parameters.

We use the sample data in MEPAVE to demonstrate the usage of AtTGen.

1. You can check the samples in data/jave_sample folder.
2. You can try this demonstration by directly running python3 playground.py.

Usage

Train on MEPAVE dataset

1. Preparing the MEPAVE dataset

Due to licensing restrictions, we cannot provide this dataset directly, please apply a license to use here, download the whole dataset and then put *.txt files in raw_data/jave folder.

2. Preprocess the data

python3 preprocess.py --dataset=jave

3. Train the model

python3 main.py --do_train --gpu_ids=0 --data_dir=./data/jave/ --ontology_vocab=attribute_vocab.json --tokenizer=char --name=jave

4. Evaluate the model

python3 main.py --do_eval --gpu_ids=0 --data_dir=./data/jave/ --ontology_vocab=attribute_vocab.json --tokenizer=char --name=jave

Train on ReCNShipNet dataset (Chinese)

python3 main.py --gpu_ids=0 --data_dir=./data/CNShipNet/ --word_vocab=word_vocab.json --ontology_vocab=attribute_vocab.json --tokenizer=chn --do_train

Train on NYT-style dataset (English, e.g. AE110k)

python3 main.py --gpu_ids=0 --data_dir=./data/nyt/ --ontology_vocab=relation_vocab.json --tokenizer=base --do_train

Reference

• OpenJERE for Tree2Sequence Loader
• JAVE for MEPAVE dataset
• ScalingUp for AE100k dataset

---

If you found this work useful, please cite it as follows:

@inproceedings{li-etal-2023-attgen,
    title = "AtTGen: Attribute Tree Generation for Real-World Attribute Joint Extraction",
    author = "Li, Yanzeng  and
      Xue, Bingcong  and
      Zhang, Ruoyu   and
      Zou, Lei",
    booktitle = "Proceedings of The 61st Annual Meeting of the Association for Computational Linguistics",
    month = july,
    year = "2023",
    address = "Toronto, Canada"
}