A sentiment is a personal feeling:
on a target (e.g., a named entity) in the text expressed with a single word or a phrase.
- Brand management
- Polling (e.g. obama)
- Purchase planning (e.g. kindle)
- Public opinion (of company and products),
- Customer satisfaction,
- Gather feedback in released products.
In the training data, to be able to find the sentiment, you can use polarity signals such as:
- Slang is the use of informal words and expressions that are not considered standard in the speaker's language or dialect but are considered acceptable in certain social settings.
If the tweet could ever appear:
- as a front\page newspaper headline
- or as a sentence in Wikipedia,
then it belongs in the neutral class. Many tweets do not have sentiment.
This approach is to use a list of positive and negative keywords. For each tweet, the number of negative keywords and positive keywords are counted. This classifier returns the polarity with the higher count.
Automatic extraction of sentiment expressions associated with given targets because the polarity of a sentiment expression is sensitive to its target.
Example with Predictable:
- With a Target Movie, Predictable is negative. Saw the movie “Friends With Benefits”. So predictable! I want my money back.
- With a stock, Predictable becomes positive.
- wiki/Naive_Bayesian_classification:classification: good, bad, neutral.
Don’t get sentimental
Everybody wants sentiment analysis. At a certain level, that’s smart. Knowing how many people are mentioning your product (or any other topic) doesn’t mean much if you don’t also know something about what they are saying.
But assessing sentiment in text is a tricky business. Humans don’t agree with one another consistently when assessing sentiment of text. In fact, even a single person asked to assess the sentiment expressed in a particular bit of text on several occasions will often give different answers. It’s hard even to make a presentation on the topic, because the audience invariably gets caught up in picking over the individual cases and debating whether the assessments are acceptable. Where's the actionable insight in that?
Instead of sentiment categories, look for something better defined and more actionable in your data. Take the example of Paypal’s Han-Sheong Lai, who uses text analytics to identify customers with intent close their accounts. Does he look for broad categories of positive and negative sentiment? No. He looks for people saying things like, “I’m going to close my account.” You can bet that makes it a lot easier to accurately assess risk, and quantity results.