Deep Learning: Sentiment Analysis

This is the semester project I did for the course "Deep Learning" taught by Daniël de Kok at the University of Tübingen in the summer semester 2017.

Using Tensorflow, I have implemented a bidirectional recurrent neural network (biRNN) that is classifying tweets according to their conveyed emotion (anger, disgust, fear, joy, sadness and surprise) leveraging information from word embeddings, entries in an emotion-annotated lexicon and character-level embeddings. My model reaches a micro-averaged F1 score of 62.81 and a macro-averaged F1 score of 52.91.

Details on the design and implementation can be found in the accompanying project plan/proposal and report (under doc).

External resources

• This project uses the (word-level) NRC Emotion Lexicon, created by Saif M. Mohammad and Peter D. Turney at the National Research Council Canada.
• It also uses the pre-trained Twitter word vectors from GloVe which are made available under the Public Domain Dedication and License v1.0 whose full text can be found at: http://www.opendatacommons.org/licenses/pddl/1.0/.

Preprocessing

The original GloVe embeddings were far too verbose for my purpose, so I stripped the unused information in a first preprocessing step, also to reduce their enormous size. Download the pre-trained Twitter word vectors from their website (I used the 200d vectors), apply get_embeddings(<filename>, 100000) from preprocessing.py to them and picklethe result into a file (cf. preprocessing.main).

Data format

The training and validation data has to be provided in a very simple tab-separated format with three columns; the first being completely irrelevant (the original data provided by the lecturer contained a tweet id), the second column containg the Tweet text and the third column containing two colons followed by the emotion class. For example:

<tweet_id>    i'm SO happy today!!! #yippie    :: joy

Usage

To run the model, use the following command:

python train.py TRAIN_DATA TEST_DATA EMBEDDINGS EMOLEX

where TRAIN_DATA is the data the model should be trained on and TEST_DATA is the data to evaluate on. EMBEDDINGS should be the file containing the preprocessed word embeddings and EMOLEX the one with the NRC Emotion Lexicon data.

For more options, refer to doc/project_report.pdf.