Sentiment analysis using sources other than the IMDB data

Question

I'm curious as to whether the training data for a sentiment analysis tool needs to be specifically geared toward the domain it's being using in. For example, the IMDB movie review data makes sense if I want to gauge sentiment on movie reviews, but would it still be okay to use if I want to gauge sentiment on what the general public thinks about a company's stock?

If I want to gauge sentiment on a company's stock, should I create my own training set that uses data from financial news/social media reaction to companies?

Answer 1

One of the most fundamental assumptions in machine learning is that the training data is 'similar' to the test data. Training makes no sense otherwise.

So the question is: How similar are reviews of, say, movies and stocks? Maybe somewhat, but not too much. Your movie-trained algorithm would certainly be able to deal with statements like 'This company is awesome' or 'Warning, do not buy this stock'.

But what about 'The stock price will explode/implode' or 'Concerning this stock, I am bearish/bullish'? The words explode, implode, bearish and bullish have probably never been expressions of sentiment in movie reviews.

Answer 2

If I understand your concern correctly, you want to know if you can perform sentiment analysis on sources other than imdb data (domains other than movie reviews that is). Short answer: YES!

Researchers in the past (even today) have tried to perform sentiment analysis on various aspects of life such as:

• Opinions of public on software/apps/products - http://ieeexplore.ieee.org/document/6912257/
• Extracting rationale from development artifacts such as issue trackers, chat messages, emails, etc. One can mine whether developers are positive or negative about a feature being introduced or the steps taken to solve them. - http://ieeexplore.ieee.org/document/8048891/
• Here is a paper which also tries to make predictions on stock markets based on sentiment analysis from twitter data: http://ieeexplore.ieee.org/document/7955659/

The challenge here would lie in the fact that you have to get some data for yourself if it already does not exists.