bcoryat / charlstm

Bidirectional Character LSTM for Sentiment Analysis - Tensorflow Implementation

Bidirectional Character LSTM for Sentiment Analysis

Requirements

• Python 2.7
• NLTK (punkt)
• Tensorflow GPU/CPU version 1.1.0 (NOTE: This code has only been tested to work on this version of Tensorflow!!)

Setup

• Download the datasets in CharLSTM/datasets
• Download the Tokenize module of NLTK using nltk.download()
• Change the PATH variable in data_utils.py and in the model files
• Install the requirements

cd CharLSTM

# Create virtual environment
virtualenv --python==/path/to/python2.7 venv

# Activate venv
source venv/bin/activate

# Install requirements
pip install -r requirements.txt

• Train the model you want with python main.py <MODEL_NAME> --train

Using a Pretrained Model

This repository provides a pretrained model for the unidirectional LSTM you can test your own sentences using:

python main.py lstm --sentences 'sentence 1' 'sentence 2' 'etc...'

Model

Results

Both models were trained for more or less 80000 iterations (~ 5 epochs) and achieved similar accuracy on a test set of 80000 tweets.

# (LSTM) Valid loss: 23.50035 -- Valid Accuracy: 0.83613
# (Bidirectional LSTM) Valid loss: 24.41145 -- Valid Accuracy: 0.82714

# Some examples...
# "cant believe i still have to write an essay..", yielded (pos/neg): 0.03065/0.96935, pred: neg
#
# "Why are you concerned with people leaking information youve declassified?" , yielded (pos/neg):
#  0.04639/0.95361, pred: neg
# 
# "Virus is going in reverse now, god my guts hurt", yielded (pos/neg): 0.09748/0.90252, pred: neg

# However, it still has problem with sarcasm:
# "his bravery?  Haha, you have to be kidding.", yielded (pos/neg): 0.73277/0.26723, pred: pos

You can read the tutorial here.