PositiveBird - use NLP to check how positive someone is on Twitter

A python web app that uses sentiment analysis on a Twitter user's tweets.

Tools:

• NLTK - for various NLP tasks, including sentiment analysis itself (with the Naive Bayes Classifier)
• twitter - Python package used to download tweets
• flask - to expose the NLP functionality via an HTTP API
• PostgreSQL + psycopg2 + SQLAlchemy - to track what users are searching for and hopefully, in the future, allow them to disagree with the analysis and give their own opinion

Data:

• Sentiment140 - has a training and a test set, both of which are used. It's pretty gigantic, but the training set is automatically, not hand-annotated. The training set is used for training, the test set is part of the test set.
• Sanders Corpus - a smaller, but hand-annotated corpus. Used as the other, bigger part of the test set.

Also, the very barebones front-end of the application can be found on the gh-pages branch of this repository.

Algorithm:

The sentiment analysis is pretty basic, and heavy inspired by sources such as this and this. Tweets in the corpus are divided into two classes, positive and negative. A Naive Bayes Classifier from NLTK is trained on the Sentiment140 training corpus, and analysis is available through a web API. The API takes a Twitter username, downloads the last 200 tweets from that user and classifies them. The result of that classification is the probability that a tweet is positive. That "positivity" is averaged over all the tweets and returned to the front-end. Then a text result is derived from that probability - either positive, neutral or negative. Neutral results occur when the probability of being positive is close to 50%.

TODO:

• Allow users to disagree with the classification, save their opinion and possibly use it to further train the classifier