Skip to content
Few-shot Text Classification with Distributional Signatures
Python Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
assets initial commit Aug 19, 2019
bin initial commit Aug 19, 2019
src initial commit Aug 19, 2019
.gitignore initial commit Aug 19, 2019
LICENSE initial commit Aug 19, 2019
README.md initial commit Aug 19, 2019

README.md

Few-shot Text Classification with Distributional Signatures

This repository contains the code and data for our ongoing work:

Few-shot Text Classification with Distributional Signatures. Yujia Bao, Menghua Wu, Shiyu Chang and Regina Barzilay.

If you find this work useful and use it on your own research, please cite our paper.

@article{bao2019distributional,
  title={Few-shot Text Classification with Distributional Signatures},
  author={Bao, Yujia and Wu, Menghua and Chang, Shiyu and Barzilay, Regina},
  journal={arXiv preprint arXiv:1908.06039},
  year={2019}
}

Overview

Our goal is to improve few-shot classification performance by learning high-quality attention from the distributional signatures of the inputs. Given a particular episode, we extract relevant statistics from the source pool and the support set. Since these statistics only roughly approximate word importance for classification, we utilize an attention generator to translate them into high-quality attention that operates over words. This generated attention provides guidance for the downstream predictor, a ridge regressor, to quickly learn from a few labeled examples.

For further details on the model and various baselines, please see src/README.md.

Data

Download

We ran experiments on a total of 6 datasets. You may download our processed data here.

Dataset Notes
20 Newsgroups (link) Processed data available. We used the 20news-18828 version, available at the link provided.
RCV1 (link) Due to the licensing agreement, we cannot release the raw data. Instead, we provide a list of document IDs and labels. You may request the dataset from the link provided.
Reuters-21578 (link) Processed data available.
Amazon reviews (link) We used a subset of the product review data. Processed data available.
HuffPost headlines (link) Processed data available.
FewRel (link) Processed data available.

Format

  • Each JSON file contains one example per line. With the exception of RCV1, each example has keys text and label. text is a list of input tokens and label is an integer, ranging from 0 to the number of classes - 1.
  • For RCV1, we are unable to distribute the original data. In place of input tokens, each example specifies a path, which corresponds to each example's file path in the data distribution.
  • Class splits for each dataset may be found in src/dataset/loader.py.

Quickstart

Run our model with default settings. By default we load data from data/.

./bin/our.sh

Scripts for other baselines may be found under bin/.

Code

src/main.py may be run with one of three modes: train, test, and finetune.

  • train trains the meta-model using episodes sampled from the training data.
  • test evaluates the current meta-model on 1000 episodes sampled from the testing data.
  • finetune trains a fully-supervised classifier on the training data and finetunes it on the support set of each episode, sampled from the testing data.

Dependencies

  • Python 3.7
  • PyTorch 1.1.0
  • numpy 1.15.4
  • torchtext 0.4.0
  • pytorch-transformers 1.1.0
  • termcolor 1.1.0
  • tqdm 4.32.2
You can’t perform that action at this time.