Project for Computational Intelligence Lab at ETH Zurich, Spring 2018

The goal of this project is to build a sentiment classifier that predicts whether a tweet text used to include a positive smiley :) or a negative smiley :(, based on the remaining text.

One-time Setup on Leonhard Cluster

The cluster is especially useful for the training of the neural networks. Login via ssh:

ssh <user>@login.leonhard.ethz.ch

Copy the script at utils/install_on_leonhard.source to your home dirctory (scp ./utils/install_on_leonhard.source <user>@login.leonhard.ethz.ch:install_on_leonhard.source)and source it. Then clone this repository git clone https://github.com/phil9987/cil_2018_text_sentiment.git.

Downloading Data

Navigate into the data/ sub-directory and download the datasets via:

curl http://www.da.inf.ethz.ch/teaching/2018/CIL/material/exercise/twitter-datasets.zip -o twitter-datasets.zip
unzip twitter-datasets.zip
mv twitter-datasets/* .

Then download the pre-trained word embeddings from the GloVe project:

curl https://nlp.stanford.edu/data/glove.twitter.27B.zip -o glove.twitter.27B.zip
unzip glove.twitter.27B.zip -d glove.twitter.27B

This should leave the data directory in the following state:

data/test_data.txt
data/train_neg_full.txt
data/train_neg.txt
data/train_pos_full.txt
data/train_pos.txt

data/glove.twitter.27B/glove.twitter.27B.25d.txt
data/glove.twitter.27B/glove.twitter.27B.50d.txt
data/glove.twitter.27B/glove.twitter.27B.100d.txt
data/glove.twitter.27B/glove.twitter.27B.200d.txt

Preparing Data

From inside the utils/ sub-directory run the pre-processing script:

cd utils
source preprocess_data.source

Preparation (for each session)

From inside the utils/ sub-directory run the activation script:

cd utils
source activate_on_leonhard.source

Baseline I: Random Forest Classifier

The code for the random forest classification is contained in the baseline/ sub-directory. It can be run via python baseline.py. On the cluster, a job should be started as follows:

bsub -B -N -n 4 -R "rusage[mem=16000,ngpus_excl_p=1]" python3 simple_rnn.py

It's worth to have a look at the documentation of scikit learn, there are many parameters which can be explored: http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html

Baseline II: Recurrent Neural Network with LSTM

The code for the recurrent neural net baseline is contained in the baseline_simple_nn/ sub-directory. It can be run as follows:

bsub -B -N -n 4 -R "rusage[mem=16000,ngpus_excl_p=1]" python3 simple_rnn.py

Our Model

The code for our own model is contained in the our_model/ sub-directory. Our approach combines a recurrent neural network with additions and tweaks to make it perform better. We utilize TensorFlow's estimator interface.

bsub -B -N -n 4 -R "rusage[mem=16000,ngpus_excl_p=1]" python3 main_v8.py

Our final, best performing model is contained in our_model/main_v8.py, but code for a couple of variations on this model are still available under our_model/archive/.