Implicit Sentiment Mining with SnowWhite

Authors: Tsvetovat, Maksim, 2042 Labs; Alex Kouznetsov,

Track: Machine Learning


SnowWhite is brand new approach to measuring sentiment in text. Instead of relying on pre-constructed lexicons of positive and negative words, SnowWhite measures secondary signs of sentiment -- specifically, a phenomenon where speakers unconsciously copy ("mirror") words and expressions of others when they feel a strong affinity.

This approach was originally tested in 2012 by attempting to predict the Republican primary election results, with a resounding success. After a number of other tests (estimating media bias during the Israel-Gaza conflict in late 2012, and analyzing response of stocks to contents of news articles), the algorithms were packaged and released as a Python module under a BSD license.

In our talk, we will outline the methods behind our novel algorithms, present results of tests and use cases, and show the attendees how to use SnowWhite to analyze texts in their own environments.