Convolutional Neural Networks for Sentence Classification

This is the implementation of Convolutional Neural Networks for Sentence Classification (Y.Kim, EMNLP 2014) on Pytorch.

Results

Below are results corresponding to all 4 models proposed in the paper for each dataset. Experiments have been done with a learning rate = 0.1 1 up to 300 100 epochs and all details are tuned to follow the settings defined in the paper.

(17.02.06) There was a bug that the best model is selected based on test acc, not on dev acc. The error has been corrected and so do results. Also note that the results can be somewhat different whenever the code is executed. Thanks! (WARNING: CV should be performed in case of MR, but only the random selection of dev(10%) and test(10%) set is done for simplicity.)

(Measure: Accuracy)

Model	Dataset	MR	TREC
Rand	Results	67.7	88.2
	Baseline	76.1	91.2
Static	Results	79.7	93.2
	Baseline	81.0	92.8
Non-static	Results	80.1	94.4
	Baseline	81.5	93.6
Multichannel	Results	79.8	93.6
	Baseline	81.1	92.2

Specification

• model.py: CNN sentnece classifier implementation proposed by Y. Kim.
• run.py: train and test a model with configs.

Development Environment

• OS: Ubuntu 16.04 LTS (64bit)
• Language: Python 3.6.2.
• GPU: GTX 1080

Requirements

This model is based on pre-trained Word2vec(GoogleNews-vectors-negative300.bin) by T.Mikolov et al. You should download this file and place it in the root folder.

Also you should follow library requirements specified in the requirements.txt.

numpy==1.12.1
gensim==2.3.0
scikit_learn==0.19.0

Execution

python run.py --help

usage: run.py [-h] [--mode MODE] [--model MODEL] [--dataset DATASET]
			  [--save_model] [--early_stopping] [--epoch EPOCH]
			  [--learning_rate LEARNING_RATE] [--gpu GPU]

-----[CNN-classifier]-----

optional arguments:
  -h, --help            show this help message and exit
  --mode MODE           train: train (with test) a model / test: test saved
						models
  --model MODEL         available models: rand, static, non-static,
						multichannel
  --dataset DATASET     available datasets: MR, TREC
  --save_model          whether saving model or not
  --early_stopping      whether to apply early stopping
  --epoch EPOCH         number of max epoch
  --learning_rate LEARNING_RATE
						learning rate
  --gpu GPU             the number of gpu to be used