A Graph-Based Neural Model for End-to-End Frame Semantic Parsing

This repository is the pytorch implementation of our paper:

A Graph-Based Neural Model for End-to-End Frame Semantic Parsing
Zhichao Lin, Yueheng Sun, Meishan Zhang
The Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021

!!!Note: we have upgrade our codebase to allennlp 2.0+ which is different from our origin implementation based on alennnlp 1.0. We find that our model can achieve further performance about 49.20 ROLE-F1 in Framenet1.5 and 49.05 ROLE-F1 in Framenet 1.7.

Installation

Clone the current repository:

git clone https://github.com/Ch4osMy7h/FramenetParser.git
cd FramenetParser

Allennlp and Other Dependencies

# Create python environment (optional)
conda create -n framenet python=3.7
source activate framenet

# Install python dependencies
pip install -r requirements.txt

Data

Request and download Framenet Dataset 1.5 and Framenet Dataset 1.7 from Website. Place them under data directory, which contatins two subdirectories called fndata-1.5 and fndata-1.7. Specially, the data should be splited into there subdirectories following FN1.5 and FN1.7. Structures are showed below:

./data
    fndata-1.5/
        train/ # containing the training xml files
        dev/
        test/
        ...
    fndata-1.7/
        train/
        dev/
        test/
        ...

To get the datasets our code based on, run the following commands:

python preprocess.py

Training & Evaluation

We use train_parser.sh script to train and evaluate all of our joint models. You can modify any configuration you want in the script.

bash train_parser.sh

Evaluation Only

All models can be evaluated using by the following commands:

allennlp evaluate your_evaluation_model \
    your_evaluation_data_file \
    --include-package framenet_parser \
    --output-file your_evaluation_output.json \

Inference on unlabled data

For predicting targets, frames and frame element on your own unlabed data (follow the file format as shown in experiments/inference/sample.json), you can use the following commands:

allennlp predict \
  --output-file your_wanted_output_file \
  --include-package framenet_parser \
  --predictor framenet_parser \
  --cuda-device 0 \
  your_trained_model \
  your_own_data.json

TODOs:

• An easy-to-use pipeline models for specific usages (will be published recently)
• An easy-to-use full joint model for extracting frame-semantic structures