While customer service is making its way to Facebook Messenger and other burgeoning platforms – a massive percentage of customer support inquiries are handled the way they have been for years: Over the phone or on chat.

As of 2015, Site Selection Group reports that there are approximately 2.2 million American call center employees in some 6,800 facilities around the US.

Call centers around the world have the same problem that other kinds of customer service centers do: They have too much data to explore. A single call center facility might generate thousands of hours of call recordings per day.

While it might be possible to quality check the calls of a single rep by manually listening to previous recordings, it isn’t realistic for a company to gain business insight from its wide array of recordings in this way. Companies want to improve the performance of individual reps, but they also want to know:

  • What are my customers reporting (or complaining) about the durability of our newest product?
  • Are we receiving more or fewer refunds after we made major revisions to our phone script?

In this article we’ll assess four Artificial Intelligence capabilities offered by the leading text analytics company, Lexalytics (our partner for this article), specifically in the area of “Natural Language Processing” (NLP) These features are: Sentiment, Entity Recognition, Categorization, and Themes. This feature set makes call center data meaningful.


Sentiment analysis is a process for answering the question: “How do they feel?” There is sentiment that expresses anger, disappointment, joy, anxiety, and more – and these feelings are important for understanding customers.

Companies used to need focus groups, or broad consumer / customer surveys to determine customer opinions – specifically asking “How do you feel about this?” Now, sentiment analysis allows companies to infer the vibe directly from these customer service interactions, getting the positive or negative straight from the customer in the moment where they are interacting with the company – without having to subject them to a post-call survey.

Sentiment not only hints at the future action of customers (i.e. angry customers may be more likely to churn or refund), it can also be matched with other data – like purchase or cancellation data – to determine clearer connections between how the tone of the interaction (positive or negative) end up affecting the bottom line. Below are some representative examples:

Airline Example:

  • An airline company that recently opened new routes to Alaska seeks to understand how the customer satisfaction of these new routes compares to it’s more established flights.
  • In this case, the airline company may draw from call center data from customers of these flights. This allows the airline company to compare the complaints and requests (and the respective frequency of both) to customers of more established routes.

Insight and Value:

  • Assume the airline realizes an overall higher incidence of negative sentiment in its complaints about in-flight service during flights to its Alaska locations. The company may investigate if this problem is due to different levels of staff performance, differences with in-flight items (i.e. food or beverages offered, headphones offered, etc), or other sources.
  • The airline may use natural language processing to detect a higher frequency of positive sentiment about flight delays, despite delay frequencies that are about equal to its other flights. The company could investigate whether this apparent reduction in complaints is due to sparse data, or due to better customer service or improved satisfaction at the gates – resulting in noticeably happier customers.

Telecom Example:

  • A telecom company may be interested in reducing churn for its high-speed internet customers in a specific geo region of high competition. With market share and revenue on the line, the company may want to double-down on the retention factors that matter most – aiming to get deeper insight into the feelings of its customers.
  • In this case, the company may want to use NLP to detect the issues and sentiments that correlated to customer retention, and those that correlate to customer churn.

Insight and Value:

  • The telecom company may realize the repeated issues of certain types are much more likely to lead to customer churn than other issues. The company may improve those most damaging customer support issues first.
  • The company might be able to match call center data with other data about the customer service process in order to glean new and unique insights. For example, the company might realize that the time between the complaint call and the arrival of a customer support vehicle to the customer’s location is a critical factor in churn. Perhaps certain time thresholds need to be upheld in order to improve retention.

Lexalytics mentions that its system is already configured to analyze directly in many languages, so additional “training” need be done only with the new entities provided by a client. Basic correlative sentiment analysis and recognition of common entities (places, prominent people, etc) are baked into the system already. Lexalytics CMO Seth Redmore explains this “configuration” as follows:

“Training requires that you have to feed a system a lot of data to get it to do what you want. Configuration can be a single line, such as ‘Don’t interpret this particular sentence as negative.’

We got this information from billions of documents where we used a set of seed terms for positive and negative… and you can extract phrases and then analyze the distance vectors from those terms. Looking at this huge volume of documents (mostly news documents from around the world), we were able to do our initial configuration.”


Sentiment is an important element of NLP business applications, but rarely is sentiment valuable in and of itself. It might be valuable for a call center manager to know the general sentiment of the center’s inbound calls in a given day or week. However, information only becomes actionable when it can be tied specifically to products, services, and other identifiable “entities.” Paul Barba, Chief Scientist at Lexalytics clarifies:

“Cuts of lumber, types of cancer, variants of a stereo model – anything that a business considers an entity — can be identified and tagged as such.”

Users looking to learn more about a specific entity (a competitor, a product, an event) can specify lists of entities ahead of time in order to track activity and sentiment around them. Lexalytics claims to also be able to detect new “entities” from data sources, potentially identifying other related nouns (for an example of this kind of entity-related NLP insight, see the Lexalytics case study with Microsoft).

Airline Example:

  • If a customer calls in referencing “SFO”, an NLP system should be trained to recognize that this is an abbreviation for an airport, and that this abbreviation represents San Francisco International Airport. Similarly, an NLP system that registers the word “Delta” or “Virgin” should recognize these terms represent individual airlines.
  • This task isn’t always as easy as detecting individual words. For example, “Alaska” can refer to a number of different entity types: “airline” or “destination state”. In situations like this, an NLP system must be trained to understand the context of a word in a sentence, in order to determine which entity to place a term into.

Telecom Example:

  • A telco company may want to know under what circumstances different telecom competitors are mentioned in phone conversations. For example, are competitors typically mentioned during refund requests, billings issues, or service requests? When are these competitors being mentioned as a comparison on service and price, as a casual mention, or as a threat of defection?


Lexalytics uses the term “themes” to describe the general “gist” of a phrase, sentence, document, or set of documents. They are automatically pulled from the text. In Paul’s words:

“Themes are lexically important noun phrases. Think of them as the ‘buzz’ from the document. They work really well when rolled up across many documents – so you can get a feel for what, exactly, are people saying.

If a reviewer Tweets ‘The bed was great but the cosmopolitans were watered down’ then Lexalytics will extract ‘hotel bar’ from ‘The cosmopolitans were watered down, despite there being no mention of ‘hotel’ or ‘bar’ in the document. Now the GM may go to the bar and share the feedback with the bar staff.”

Themes play a role in determining patterns of communication – even when the same words aren’t used to describe the experience. Lexalytics claims that themes are extracted automatically through text mining of the raw data from call recordings.

Airline Example:

  • An airline looking to understand customer sentiment around wait times in a specific geo regions. While searching for sentiment only around the words “wait,” “waited,” and “waiting” will probably derive some insight, it might not be complete. A “themes” feature might unearth sentiment and common statements related to terms like “standing around” or “being held up” or “sitting forever,” providing a more complete perspective on the theme of “waiting.”


 Paul Barba explains the concept of NLP “categories” this way:

“Categories are the other side of the ‘determining context’ coin from Themes. Themes are extracted completely automatically, where categories need to be configured ahead of time. This is useful for sorting content into buckets that are useful and relevant to a business. For example, with a retail establishment, they might be interested in categories such as staff, location, parking, stock availability, lighting, pricing, etc.”

Categories are manually entered “buckets” which can be used to sort the otherwise disorganized stream of support calls. Paul clarifies how this technology works at Lexalytics.

“We do have automatic categories, and these are very high-level buckets for you to use to get a preliminary view of the content. With several different ways of categorizing content, from search queries to machine learning classifiers to Wikipedia-based categories, we provide the text analytics tools necessary to segment content exactly the way that is most relevant to any business.

Basically, categories serve to answer the following questions: ‘What is this content about?’ or ‘What concepts does this content mention?’”

Airline Example:

  • An airline interested in learning more about why and when its customers change flights might keep an active pulse on “flight change requests” as a category of support calls using NLP. This would allow the airline to analyze the sentiment, entities, and terms in just “flight change request” calls in order to learn more about this request and better serve their customers.

Telecom Example:

  • A telecom company might be interested in determining the percent of its calls categorized as “service requests” or “hardware problems” in order to determine when these categories of customer service requests spike. The company can then later explore the reason for these spikes and aim to fix the root concerns (possibly related to weather, hardware durability, or other factors).

Categories must be determined upfront, but they also must be trained with initial data from a user company. Many representative examples of a “refund call” or “flight change call” or a “service request call” must be provided to an NLP engine in order to accurately detect the patterns that encapsulate that category.

More from Lexalytics

This article was sponsored by and created in partnership with Lexalytics, a text analysis service provider. Lexalytics analyses text (from call center transcripts, social media feeds, online media, and more) in order to surface insight for their business clients.


This article was written in partnership with Lexalytics. For more information about content and promotional partnerships with TechEmergence, visit the TechEmergence Partnerships page.

Header image credit: AllStaff