Text Analytics as used in Contact Centers

     

Evolution of Unstructured Data in Contact Centers

Having spent more than 15 years in contact centers that been the focal point of customer interactions of all kind – signing up for a credit card, subscribing for an energy connection to getting your computer issues solved through phone or an online channel to requesting for cancellation from different subscription businesses – one thing I see common amongst the half a dozen organizations that I have worked for. All of them handle in excess of 10 million transactions per month and there is a huge opportunity to understand customer behavior during different stages of the product lifecycle. All these organizations handle structured data around these interactions extremely well. There are many meaningful reports that are churned out of this data from different systems –Switch, Ticketing System, Surveys, Order Management System etc..

Around 2004-05, most US clients started looking at Philippines as an option compared to Indian cities. The fact that the Americans ruled the South East Asian nation had definitely had a role to play in its present culture which is so much “Follow the Yankees”.  A little later, around 2007, webchat started playing a huge role in the contact mix of CARE interactions. A lot of Fortune 500 organizations started their digitization strategy with chat and then virtual avatar aka intelligent self-service strategies.  This ensured the availability of the text transcript [chat] in addition to all the data around it – like which page did the customer come from, what keyword did he use, what IP address did he use, which region is that IP address assigned to etc…This started the revolution of the usage of unstructured data analysis in contact centers.

Text Analytics defined…

Text Mining refers to the process of deriving high-quality information from text. This is achieved by transposing words and phrases in unstructured data into numerical values which can then be linked with structured data in a database and analyzed with traditional data mining techniques. The application of text mining techniques to solve business problems is called text analytics. A lot of websites will refer to both “text mining” and “text analytics” interchangeably.

Text Mining Architecture

The text mining process can be based on either linguistic or statistical or a combination of both approaches. Gartner breaks the entire ecosystem into 3 stages –

1.       Data Collection [Text acquisition and preparation]

2.       Data Processing [processing and analysis] and

3.       Data Representation [end output]

A High Level Technical Architecture of a Text Analytics System is provided in the framework below

Source: Gartner (August 2014)

More details of the activities and outputs in each stage are provided in the table below.

In this paper, we will discuss the data processing aspect and less on the 1 -data collection and 3- data representation side.

2.      Data Processing

Machine Learning for Unstructured Data

Unstructured data offers terabytes of information. That presents a huge opportunity as well as a big challenge to synthesize meaningful information. Machine learning provides a huge opportunity to synthesize the volume of data that is available and is one of the most common approaches to deal with large volumes of data. Machine learning is a scientific discipline that explores the construction and study of algorithms that can learn from data. Such algorithms operate by building a model based on inputs and using that to make predictions or decisions, rather than following only explicitly programmed instructions. Machine Learning can be considered a subfield of Computer Science and Statistics. In 1959, Arthur Samuel defined machine learning as a "Field of study that gives computers the ability to learn without being explicitly programmed"

Most researchers suggest that Linguistic approaches are best used for text analysis instead of machine learning techniques. Here is an example of how a leading text mining software categorizes some of the text available in a survey based on the standard library.

“Get rid of the piece of crap automated system you operate and give me an option to go to a knowledgeable human directly.” – categorized as an Agent Knowledge issue by the text mining software whereas this is a complaint about the IVR

“The ease of sending & receiving money without worrying about my details falling into the wrong hands” - – categorized as an Agent “Easy to Understand” by the text mining software whereas this is more about the process of transferring money

Linguistic Approaches as an alternative for Machine Learning Techniques

Natural language text is not a medium readily understood by computer systems, in contrast to the neatly arrange rows and columns in a database. This is the primary reason that text analytics has had such a long gestation before it could be usefully employed in a business arena. Our experience suggests it is best to use deep linguistic analysis to generate a stable data dictionary with high coverage and accuracy and then use machine learning to automate these. In the first step, which is to identify words A combination of these approaches provide the best possible when the data volumes are high.  Coverage and Accuracy around 70% is a tall order in such cases whereas business expects coverage and accuracy of 90%+.

The data preparation aspect brings the need for text tagging. Text Tagging is done in 2 different ways as of now

1.       Human Tagging : This is non-scalable and subject to differences between different resources

2.       Auto Tagging : Uses an automated system based on unsupervised machine learning techniques

How do you combine the best of human tagging and auto tagging to generate a cost-effective solution?

1.       Use the keyword extraction feature of the text analytics tool to look at the keywords or phrases that show the # of occurrences

2.       Unsupervised category-document correlation generates numerical scores for each document relative to a set of interesting topics.

3.       Have a human tagger with process knowledge of the interactions use these keywords and phrases to build the initial set of data dictionaries

4.       Regression with supervised machine learning generates numerical probability that a document belongs to one of many 'classes' established from ground-truth.

CONTEXT is KING

While these are examples of incorrect text categorization, it sends a larger message that CONTEXT is more important in Unstructured Data Analytics. Definition of Context is a challenge that every linguistic analyst is set to conquer. A good linguistic approach uses the following approaches

1.       A data dictionary that works with keywords does not provide you with an accurate insight as languages are more complex e.g. social security number being picked up for security topics. Look for phrases. The expression "kick the bucket" may not make sense to someone unfamiliar with American idioms.

2.       Relation Extraction when sentences follow subject, action and object order instead of a bag of words approach.

3.       Distance is a key factor in dialogues as most context in a response  is appropriate only in the immediate distance instead of the entire length of the chat text

4.       Modifiers and Negations– It is important to use modifiers and Negations. Adverbs and Adjectives are used to qualify a verb and/or noun. There is a big difference between “good”, “very good” and “not so good”

5.       Sentiment Analysis – While providing a sentiment for the keyword is something most text analysis engines can achieve, providing a sentiment score for each document based on the modifier and the negation for the whole document

6.        Visualization and Search Access – It is important to allow the user to visualize the analysis as well as to provide search functionality for the business user to understand the context better

How is text mining combined with predictive analytics effective in Customer Experience Management?

Text Mining of the verbatim in surveys and the chat transcript allows the organization to identify contextual variables that add more details to the existing structured data set for the transaction. In addition, there is an opportunity to provide a sentiment score for the interaction. This score and the other variables allow prediction of the NPS for 100% of the interactions – a huge improvement from the 5-20% response rate that is typically available. A single view of all these variables [structured, unstructured and predicted] for the leadership and alerts for poor predicted NPS allows the Organization to close the loop with not just the detractors who responded to the survey but also those who in other cases would have just switched.

Build versus Buy

Like every technology solution, there is an option for every user to buy, build or partner with a provider for re-use.  Some of the criteria that determine this are as follows

1.      Customization needed for different engagements and different data sources– If you are trying to set the data dictionaries once and have minimal changes and you do not need to constantly add new data sources

2.      Integration to other systems and Support needed – If you are looking to integrate this to your social media listening tool or your customer experience management tool in the contact center or a Visual analytics tool for multiple users

3.      Skillset available in the organization

4.      Cost

Text Analytics Providers

In case you have decided to buy licenses, some of the top text analytics solutions are provided by the following providers - SAS Text Analytics; IBM Text Analytics; Lexalytics; Clarabridge and Attensity

Glossary

1.      IP - Internet Protocol

2.      IVR - Interactive Voice Response Unit

3.      NPS - Net Promoter Score

4.      ASAT - Agent Satisfaction Score

5.      FCR - First Call Resolution

Sources: Wikipedia.com, AllAnalytics.com, Gartner.com