A BiLSTM+GCN Encoder for Target-oriented Opinion Words Extraction

This repo contains the official PyTorch implementation on the best performing model (BiLSTM+GCN) for paper An Empirical Study on Leveraging Position Embeddings for Target-oriented Opinion Words Extraction published in the proceedings at EMNLP 2021.

The TOWE datasets: Details on the TOWE Datasets can be found in (Fan et al., 2019).

Requirements

• Python 3 (tested on 3.6.13)
• PyTorch (tested on 1.9.0)

Preparation

First, download and unzip 300-dim GloVe vectors from the Stanford website into the directory data/glove/

wget https://nlp.stanford.edu/data/glove.840B.300d.zip
unzip glove.840B.300d.zip -d data/glove/
rm glove.840B.300d.zip

For each dataset (e.g., 14lap), prepare vocabulary and initial word vectors with:

python prepare_vocab.py --dataset 14lap --wv_file glove.840B.300d.txt

This will write vocab.pkl and embedding.npy into the dir data/14lap.

Training and Evaluation

Train and evaluate a BiLSTM+GCN encoder with:

python train.py --dataset 14lap --gcn_layers 1 --save_dir best_model_log

This will train a BiLSTM+GCN encoder with a single GCN layer.

The argument --gcn_layers takes an integer for the encoder

Training a BiLSTM encoder has the corresponding gcn_layers: 0.

Alternative Training and Evaluation

Train and evaluate all datasets for different number of layers

python run1.py

Citation

If this work is beneficial to you, please cite as:

@inproceedings{mensah-etal-2021-empirical,
    title = "An Empirical Study on Leveraging Position Embeddings for Target-oriented Opinion Words Extraction",
    author = "Mensah, Samuel  and
      Sun, Kai  and
      Aletras, Nikolaos",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.722",
    pages = "9174--9179",
}