Sentiment Analysis methods classify data, such as customer reviews, by negative or positive sentiment based on an individual’s opinions or attitudes toward any object (e.g., products, services, etc.). Typically, these methods can find evaluative terms or phrases, like “great” or “bad” or “very tasty”, around an object of interest, and then classify them into negative or positive sentiment, and sometimes also neutral sentiment.
However, classifying an evaluation as just positive, negative, or neutral can be limiting. In classification approaches to sentiment, we need to account for certain nuances to evaluative language, such as the strength of sentiment. For example, a human reader can identify the difference between the following two positive statements:
“That hamburger was okay”
“That hamburger was excellent”
We know that a customer who describes a product or service as “okay” may be a satisfied customer overall; however, this person may also have mixed feelings about a product or service that need to be addressed. On the other hand, a customer who describes a hamburger as “excellent” is probably fully satisfied.
If a computer only classifies these reviews as positive or negative, it only sees two equally positive words and two equally positive reviews. It does not know how to handle such differences unless you teach it to do so.
Many companies already try to pinpoint these strong reviews with customer surveys in which the customer is prompted to rate products and services on a scale from “strongly dislike” to “strongly like”. Imagine if you could automatically convert candid, text-only online customer reviews into the same scales. In this sense, strong positive or negative sentiment can represent a point on the traditional numerical scale.
As in customer surveys, the strength of an evaluation can be represented by assigning sentiment scores or degrees to groups of positive or negative evaluative terms or phrases. For example, “excellent” is a much stronger positive sentiment than “okay”, so it is assigned a higher positive sentiment score. Similarly, if the customer described the hamburger as “terrible” (or with curse words), it is likely to be strongly negative.
By classifying evaluative terms by strength and polarity, you no longer need to rely on developing costly and time-consuming customer surveys to locate very satisfied and very unsatisfied customers. Instead, you can harness data from review sites or social media to automatically detect not only how positively or negatively customers feel about a product or service, but also how strongly they feel about it.
Why is sentiment degree-setting important?
By numerically representing the degree of a sentiment, it is possible to create more nuanced visualizations and efficient models for tracking sentiment. Below are a few case examples of how illustrating sentiment strength can lead to better business decisions.
Tracking Sentiment Growth Trends. The numerical values assigned to sentiments are easily translated to trend analysis models. For example, by summing the degrees that represent real evaluations, such as okay and awesome, it is possible to visualize changes in a product’s evaluation. Below is a visualization of the positive and negative sentiment growth of a company’s products. It is easily noticeable that the company’s “pricing” has seen the strongest improvement in the eyes of their customers, whereas the “quality” of their products has seen only marginal sentiment growth.
Identifying Sentiment Extremes. It is important to know which customers are extremely satisfied or unsatisfied in comparison to those with slightly more neutral attitudes. For example, it is possible to sum the sentiment degrees and plot the most satisfied and unsatisfied customer bases on a GIS map. In the map below, green represents a satisfied customer base, and red, an unsatisfied one. Extremely negative sentiment (dark red) represents a customer base that has the most immediate need for improvement in its business relationship; in this example, Tunisia and Indonesia. Because this sentiment model was built with negative and positive terms associated with products and services, this company can look up the sentiments contributing to a country’s score and pinpoint exactly what needs improvement, such as pricing, availability of products, or customer service responsiveness.
Identifying Neutral Populations. As a company, you may still want to focus on reaching the customer bases with less extreme sentiment, represented in yellow on the GIS map above. If a customer is reporting a less extreme evaluation of a product, it is still important to reach out to these customers so your company can positively influence customer opinions and efficiently improve products and services as a result. For instance, Brazil seems fairly neutral about its business relationship with this company, so it would be recommended to assess which products and services are weighing down its overall assessment.