Understand Sentiment Analysis in short article

We’ve said that sentiment analysis takes a text document as input and returns a representation of a sentiment as output.

There is little to say about the input. This is simply the text content of the book/comment/customer survey/email/news article/product review/tweet or other type of document that you would like to analyze.

Now, let’s turn to the output.

Binary sentiment analysis

Binary sentiment analysis, the simplest case, asks the following question: “Is the opinion expressed in the text document positive or negative?”

Here, the output is either a probability or a score.

Let’s consider probabilities first.

A high probability indicates that the given text is likely to express a positive opinion. For example, an output of 0.9 indicates a 90% probability that the expressed opinion is positive.

Conversely, a low probability indicates that the given text is likely to be an expression of a negative view. For example, an output of 0.1 indicates a 10% probability that the opinion is positive or, put differently, a 90% probability of a negative opinion.

Alternatively, the prediction of the sentiment can be expressed as a score. There are no surprises here, either. Positive scores indicate positive sentiments. Negative scores indicate negative sentiments. Greater absolute values indicate a stronger sentiments.

Multi-class sentiment analysis

In multi-class sentiment analysis, the output is a probability distribution over n possible classes or, in other words, a list of n non-negative numbers that add up to 1. To categorize the opinion, you can pick the classes associated with the highest probability.

Binary sentiment analysis is a special case of multi-class sentiment analysis with n=2.

A simple extension of binary sentiment analysis is to introduce the class of neutral opinions. In that case, the output is a probability distribution over three classes (n = 3). The three probabilities could be ordered as follows: negative probability, neutral probability and positive probability.

The prediction for a review that focuses on technical details could, for example, have a distribution similar to this one: [ 0.1, 0.85, 0.05 ].

Now, suppose that a customer publishes a mixed review, listing both positive and negative aspects of a product. The output of the 3-class sentiment analyzer could resemble the following distribution: [ 0.4, 0.1, 0.5 ].

One way to deal with this is to include a mixed category. Suppose we order the four categories as follows: negative, neutral, mixed, positive. The output in the mixed product review example might then look something like this: [ 0.15, 0.05, 0.7, 0.1 ].