Dynamic Semantic Matching And Aggregation Network for Few-shot Intent Detection

This repository provides PyTorch implementation for the paper Dynamic Semantic Matching and Aggregation Network for Few-shot Intent Detection (Findings of EMNLP'2020)

Requirements

Python 3.6.2
Numpy
Pandas
Pytorch 1.0.1
Scikit-learn 0.21.1

Dataset

We conduct the split on NLUE and SNIPS dataset in dataset directory. Please take a look at our paper for details of the split.

Usage

Please obtain and put the pre-trained FastText embedding in our fasttext directory of the repository (named as vectors-en.txt). Otherwise, create your own FastText embedding directory and update the argument in Configuration below.

Configuration

• --ckpt_dir: Saved directory for checkpoint
• --eps: Evaluate as nonepisodic or episodic procedure (i.e. eps or noneps)
• --num_eps: Number of episodes used for episodic training and/or evaluation
• --dataset: Choose dataset to train/evaluate (i.e. SNIPS/ NLUE)
• --num_run: number of runs (only for SNIPS)
• --num_fold: number of KFold counting from 1 to 10 (only for NLUE)
• --src: Source data used for training (i.e. seen)
• --tgt: Data used for evaluation (i.e. novel or joint)
• --num_samples_per_class: K in C-way K-shot
• --num_class: C class in C-way K-shot
• --num_query_per_class: num query per class (Q)
• --num_test_class: Number of classes used for evaluation (i.e. C for episodic, #total classes in joint/novel space)
• --fasttext_path: FastText pretrained embedding file location

Regularization hyperparameters

• --self_attn_loss: coefficient for Self-attention Regularization
• --uniform_loss: coefficient for Head Uniform Regularization
• --same_intent_loss: coefficient for Head Distribution Regularization

Running

SNIPS_Dataset

FSL (episode-nonepisode)

python main.py --dataset='SNIPS' --tgt='novel' --eps='eps' --num_class=2 --num_test_class=2

python main.py --dataset='SNIPS' --tgt='novel' --eps ='noneps' --num_class=2 --num_test_class=2

GFSL (episode-nonepisode)

python main.py --dataset='SNIPS' --tgt='joint' --eps='eps' --num_class=2 --num_test_class=2

python main.py --dataset='SNIPS' --tgt='joint' --eps ='noneps' --num_class=2 --num_test_class=7

NLUE Dataset

FSL (episode-nonepisode)

python main.py --dataset='NLUE' --tgt='novel' --eps='eps' --num_class=5 --num_test_class=5

python main.py --dataset='NLUE' --tgt='novel' --eps ='noneps' --num_class=5 --num_test_class=16

GFSL (episode-nonepisode)

python main.py --dataset='NLUE' --tgt='joint' --eps='eps' --num_class=5 --num_test_class=5

python main.py --dataset='NLUE' --tgt='joint' --eps ='noneps' --num_class=5 --num_test_class=64

Citation

If you use our ideas, code or dataset, please cite the following paper:

@inproceedings{nguyen-etal-2020-dynamic,
    title = "Dynamic Semantic Matching and Aggregation Network for Few-shot Intent Detection",
    author = "Nguyen, Hoang  and
      Zhang, Chenwei  and
      Xia, Congying  and
      Yu, Philip",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.108",
    pages = "1209--1218",
}

Acknowledgement

https://github.com/ZhixiuYe/MLMAN
https://github.com/galsang/BIMPM-pytorch