Transfer Learning for Text Classification with Tensorflow

Tensorflow implementation of Semi-supervised Sequence Learning(https://arxiv.org/abs/1511.01432).

Auto-encoder or language model is used as a pre-trained model to initialize LSTM text classification model.

SA-LSTM: Use auto-encoder as a pre-trained model.
LM-LSTM: Use language model as a pre-trained model.

Requirements

Python 3
Tensorflow
pip install -r requirements.txt

Usage

DBpedia dataset is used for pre-training and training.

Pre-train auto encoder or language model

$ python pre_train.py --model="<MODEL>"

(<Model>: auto_encoder | language_model)

Train LSTM text classification model

$ python train.py --pre_trained="<MODEL>"

(<Model>: none | auto_encoder | language_model)

Experimental Results

Orange lines: LSTM
Blue lines: SA-LSTM
Red lines: LM-LSTM

Loss

Accuracy