Sentiment Analysis Helps Get the Most out of Customer Surveys

Customer Surveys Contain Critical Feedback

For companies, it is critical to gather and analyze as much customer feedback as possible.  This is true of almost every industry and especially so of consumer-oriented industries such as hospitality, consumer products, airlines, retailers, entertainment, etc.  It’s not just in the commercial world that customer feedback is important.  Public agencies also gather feedback on how their services are being received, how their intended audience learned about them, and so on.

A traditional way of collecting customer feedback is through surveys.  Surveys are much more targeted than gathering data from social media and they remain essential for an organization to understand how it and its products and services are doing, how they fare against competitors, and what additional features, offerings, or services customers may be looking for.  Surveys can provide a wealth of information on all aspects of customers’ experience including such issues as product/service quality, performance, functionality, price, customer support, etc.

Handling the Unstructured Responses in a Survey Is Challenging

In addition to having customers provide a numerical ranking in response to survey questions (e.g., from 0 = Disappointed to 5 = Delighted, with intermediate rankings in between), there are often free text questions (e.g., “Please comment on how we could make your hotel stay more enjoyable”). Free text, known in our field as “unstructured data,” will frequently contain insightful and actionable information and is often expressed in the form of fine-grained sentiment about specific features of a product or service (e.g., “I expected a better coffee selection”, “towels smelled moldy”).  While it’s easy to aggregate numerical rankings, it is much harder to analyze and track at scale and in a timely fashion the sentiments contained in unstructured data.

Sentiment Analysis Automates the Process of Analyzing the Unstructured Data in Surveys

Rather than having staff laboriously read through such survey answers and coding them numerically, a promising technology called Sentiment Analysis can process large amounts of unstructured data very quickly and automatically produce a quantitative understanding of the sentiment contained in it.

Using Sentiment Analysis technology allows survey designers to rely more on free-text questions to collect highly valuable nuggets of insight that a strictly numerical ranking survey could not capture.  Sentiment Analysis also achieves a higher level of consistency than when humans read the responses.

How Sentiment Analysis Works

There are different approaches represented in the various Sentiment Analysis products on the market today. The simplest approach involves simply looking for key words like “great” or “terrible” and then characterizing the sentiment of the document accordingly.  A much better approach is to find sentiment-bearing words or phrases and then connect them to the specific item they are referring to in the context. This is known as Entity-Based Sentiment Analysis and allows a much more fine-grained analysis.

How Entity-Based Sentiment Analysis Can Be Used to Analyze Customer Comments

Here’s a typical customer comment you might see in the unstructured portion of a survey:

“The product set-up was easy and quick, but the package arrived damaged and the spare lithium batteries were slightly dented”

There are four sentiments being expressed here, two positive and two negative. The output of an advanced Sentiment Analysis product would look like this:

• sentiment polarity: positive
• sentiment predicate: easy
• sentiment object: product set-up

 

• sentiment polarity: positive
• sentiment predicate: quick
• sentiment object: product set-up

 

• sentiment polarity: negative
• sentiment predicate: damaged
• sentiment object: package

 

• sentiment polarity: negative
• sentiment predicate: dented
• sentiment object: lithium battery

The four unstructured sentiments expressed in the customer comment have been parsed (to identify sentiments and objects) and normalized (e.g., “slightly dented” -> “dented”, “the spare lithium batteries” -> “lithium battery”).  The survey’s unstructured data is thus transformed into structured data that can be aggregated, quantified, and tracked over a very large set of such comments as well as over time to produce the actionable insights an organization needs. Surely if a significant number of surveys suddenly mention damaged packaging and dented goods, a company would want to be alerted right away in order to address any shipping issues promptly.

In sum, understanding what the customer is saying is very difficult in a crowded commercial space.  Sentiment Analysis is necessary to automate the otherwise painful, slow process of wringing true value out of customer surveys.