Sentiment Analysis on Twitter

The Problem

Given a tweet (that contains some text), estimate the sentiment (negative or positive) of the tweeter.

Training, Development, and Test Datasets

Some folks at Stanford spent more than a year doing research on sentiment analysis on twitter. They published a paper [here] (http://cs.stanford.edu/people/alecmgo/papers/TwitterDistantSupervision09.pdf) and released both their training and test sets, which we used throughout our project.

The training set has 1,600,000 tweets marked positive/negative, while the test set has 498 tweets. We extracted a development set of 500 tweets from the original training set to use in adjusting various parameters for both types of classifiers.

Methods Used

• Naive Bayes Classifier: Using the "best" parameters on our naive bayes classifier, we achieved an accuracy of 83.01%. It turns out that our Naive Bayes classifier performs better than that of Alec Go's. The Naive Bayes classifier is implemented in naivebayesclassifier.py and the Naive Bayes evaluator (which measures the effectiveness of the classifier) is implemented in naivebayesevaluator.py.

  Run python naivebayesevaluator.py -g 1 1,2 to see the accuracy on the test set.

  Run python naivebayesevaluator.py -h to see all options for the Naive Bayes Evaluator. For more details, see documentation in the evaluator file.

• Maximum Entropy Classifier:
  The input to an instance of the Maximum Entropy Evaluator is made up of four parameters:

  • Number of tweets to train on (filesubset)
  • Minimum number of occurences a feature must have appeared to be included as a feature (min_occurences)
  • The number of iterations to run GIS (max_iter)
  • What n-grams to use (grams, a list)

  These parameters can be tweaked in maxentevaluator. Generally, we noted the intuitive trend that using more data gave better results. As a result, to get the best results, we would use a large subset of tweets and a small threshold level of feature occurences (i.e. including as many n-grams as possible). For practical purposes of demonstration, the parameters are set to lower values which only take a few minutes to run.

  In order to use a better model (which achieves 76% on the test set), we've included a pickled model which was trained using 4000 tweets, unigrams and bigrams, and a threshold of 3. When a Maximum Entropy Classifier is trained, the resulting model is pickled to maxentpickles. While we have not included all the pickled models, the file maxent_4000_3_2.dat is included and can be run via maxentevaluator's runFromPickle method.

Web Interface

Our models were trained and evaluated on data accumulated in 2006. We thought it might be useful (and cool) to evaluate our models (Naive Bayes and Max Entropy) on more recent tweets using the Twitter real-time API. So we setup a python Tornado server with a Graphical User Interface (built using HTML + CSS + Javascript) to grab tweets, perform sentiment analysis on these tweets, and display the results in an intuitive manner.

If you want to run the web server, you need tornado web (which can be easily intalled via pip or easy_install. Use python app.py to startup the server. The file first loads both classifiers (takes about 2 minutes to load the models). Wait till you see the messages: Model retrieved from 'model1-2.dat' and Max ent model built. By default, it runs on port 8888. To see the server in action, using your favorite browser go to http://localhost:8888/. Pick what model you want to evaluate the tweet on, and Search.

MIT Open Source License

Copyright © 2012 Daniel Alabi, Nick Jones

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.