Debunking NLP: Detecting Actionable Sentences

By: Ellen Loeshelle

November 17, 2017

Clarabridge Analytics
Customer Effort
Artificial Intelligence
Machine Learning
Natural Language Processing
Natural Language Understanding

This blog post is part of our Debunking Natural Language Processing (NLP) series. Throughout this series, Ellen will highlight several features that help Clarabridge users go beyond simple topic analysis. This series will show you how new types of analysis aren’t so farfetched after all!

Throughout college, our refrigerator was littered with Magnetic Poetry magnets from the various languages that my roommates and I spoke or were studying. English, Spanish and Yiddish nouns, verbs, adjectives plus corresponding function words such as prepositions, conjunctions and pronouns, as well as a variety of profanity, allowed us to imagine novel phrases that admittedly often had somewhat NSFW intent. Being able to reorganize words tactilely propels us to reimagine their meaning and how they could be reorganized to achieve new intent. While we are fixated briefly on each individual magnet, it is the complete clauses or sentences which yield paradoxical, nonsensical or absurdly amusing meanings that cause us to laugh until tears.

In my past three blog posts in this series, we focused our discussion on the value in extracting “Named Entities,” or proper nouns such as brands, locations and persons, from customer feedback data. These words often carry more weight than other nouns as they refer to a specific entity that reveals preferences and biases which could affect a customer’s perception or experience. However, like the Magnetic Poetry example, individual words are so much more meaningful when in the context of a sentence.

We normally don’t spend a lot of time thinking about the “type” of sentence that we are uttering. Each sentence is just a piece of a larger dialogue used to convey our message. But, when analyzing customer feedback, being able to isolate certain kinds of sentences actually lends substantial power. Compare “Your website sucks!” and “Your website would be so much easier to use if the chat box didn’t cover up the login area!” While we might be drawn to the obvious negativity in the first sentence, it is the second one which we would deem as actionable. It offers an explicit suggestion that actually unlocks valuable information which allows us to identify specific pain points and to design customer-centric solutions.

Clarabridge uses semantic analysis strategies to identify 13 different kinds of sentences specifically for CX analytics. Together, the three actionable types (Suggestions, Requests and Cries for Help) and the ten non-actionable types (No Comment, Don’t Know, Everything, Yes, List, Cross-Reference, Generic Praise, Generic Apathy, Hello/Goodbye and Thanks) give Clarabridge users the power to segment their data in ways that go far beyond traditional topic analysis. As some users struggled to identify these kinds of feedback using keyword searching, we realized that the variation in language was too great to solely rely on this strategy. So we hit the drawing board and returned with a machine learning based solution that leverages all of the linguistic goodness that underlies the Clarabridge NLP to identify sentence types rather than simply keywords and phrases.

How is the Sentence Type attribute best leveraged? Here are the top use cases for these features:

1- Isolate Suggestions and or Requests

Let your customers guide you to success by isolating their suggestions/requests throughout your dataset. Analyze the top topics of suggestions to understand where your business could improve; track KPIs for this customer group to determine how any product or service deficiencies are affecting your business. Look at topics of requests (questions) to determine where you may need to amp up your messaging or clarify content on your website.


2- Filter Out Noise

Hearing the customer’s voice can be challenging when it’s swallowed up in a sea of non-actionable messages. Use our various non-actionable sentence types to filter out the content that you don’t want to see. This step helps to part the seas, so to speak, allowing you to hear the feedback that you were missing.

3- Route Cries for Help

Don’t miss out on an opportunity to help your customer! When customers ask for a return contact or leave their phone number/email address in their response, they are pleading for assistance. Use the Cries for Help sentence type to automatically trigger an alert and a case within our Case Management system so that it gets routed to the agent that can help them best.

With increasing pressure on CX teams to deliver value and customer-centric solutions, there’s no time in the day to meander around looking for insights. Topic reports lack the oomph needed to truly change a business for the better. By analyzing the kinds of sentences in their data, analysts can better understand their customers needs and wishes and cut through all of the noise with a few swift clicks. It’s no longer sufficient just to know what your customers are talking about; we need to also understand how they’re talking about it!

To read the previous blog posts in this series, please visit:

Debunking NLP: Introduction

Debunking NLP: Translation

Debunking NLP: Named Entities

Debunking NLP: Detecting Products, Brands, Companies and Industries

Debunking NLP: Detecting People, Phone Numbers and Email Addresses