Detecting the presence of depressive characteristics in tweets.

Project Overview

Data Collection and preprocessing

Notebook:tweepy.ipynb -Used social media data from the Twitter API .Using the Tweepy python library I was able to extract close to 2000 tweets using relevant keywords. Used Natural Language Toolkit to preprocess and clean the tweets.

Automatic Labelling.

Notebook:tweet_emotions.ipynb Attached labels to the tweets ranging from (1-5) where 1 signified least likely to be depressed while 5 signified most likely to be depressed. Used 3 lexicons namely Affin,Vader and NRC to generate the scores . Performed a weighted average of the scores to correctly classify the tweets. From the NRC emotion lexicon, emotions which are common in a depressed individual like sadness and fear were assigned greater weights while emotions like Joy and Trust were penalised and assigned smaller weights.

Training the Neural Network

model.ipynb Used Word2Vec to prepare a word embedding which was then fed to a simple LSTM network.

References:

• Detecting the magnitude of depression in Twitter users
• You are what you tweet