SWEM (Simple Word-Embedding-based Models)

This repository contains source code necessary to reproduce the results presented in the following paper:

• Baseline Needs More Love: On Simple Word-Embedding-Based Models and Associated Pooling Mechanisms (ACL 2018)

This project is maintained by Dinghan Shen. Feel free to contact [email protected] for any relevant issues.

Prerequisite:

• CUDA, cudnn
• Python 2.7
• Tensorflow (version >1.0). We used tensorflow 1.5.
• Run: pip install -r requirements.txt to install requirements

Data:

• For convenience, we provide pre-processed versions for the following datasets: DBpedia, SNLI, Yahoo. Data are prepared in pickle format, and each .p file has the same fields in the same order:

  • train_text, val_text, test_text, train_label, val_label, test_label, dictionary(wordtoix), reverse dictionary(ixtoword)

• These .p files can be downloaded from the links below. After downloading, you can put them into a data folder:

  • Ontology classification: DBpedia (591MB)

  • Natural language inference: SNLI (101MB), SNLI-word-embeddings (129MB)

  • Topic categorization: Yahoo (1.7GB)

Run

• Run: python eval_dbpedia_emb.py for ontology classification on the DBpedia dataset

• Run: python eval_snli_emb.py for natural language inference on the SNLI dataset

• Run: python eval_yahoo_emb.py for topic categorization on the Yahoo! Answer dataset

• Options: options can be made by changing option class in any of the above three files:

• opt.emb_size: number of word embedding dimensions.
• opt.drop_rate: the keep rate of dropout layer.
• opt.lr: learning rate.
• opt.batch_size: number of batch size.
• opt.H_dis: the dimension of last hidden layer.

• On a K80 GPU machine, training roughly takes about 3 minutes each epoch and 5 epochs for Debpedia to converge, 50 seconds each epoch and 20 epochs for SNLI, and 4 minutes each epoch and 5 epochs for the Yahoo dataset.

Subspace Training & Intrinsic Dimension

To measure the intrinsic dimension of word-embedding-based text classification tasks, we compare SWEM and CNNs via subspace training in Section 5.1 of the paper.

Please follow the instructions in folder intrinsic_dimension to reproduce the results.

Citation

Please cite our ACL paper in your publications if it helps your research:

@inproceedings{Shen2018Baseline, 
title={Baseline Needs More Love: On Simple Word-Embedding-Based Models and Associated Pooling Mechanisms}, 
author={Shen, Dinghan and Wang, Guoyin and Wang, Wenlin and Renqiang Min, Martin and Su, Qinliang and Zhang, Yizhe and Li, Chunyuan and Henao, Ricardo and Carin, Lawrence}, 
booktitle={ACL}, 
year={2018} 
}