Database : Text Mining in Marketing

An analyst’s perspective on reading lips … and prompting action

In the world of marketing, text mining has gained greater awareness and favor as social media, inbound e-mail marketing, filtering and other "digitized" communication and documents in free text have flourished—and as the cost and quality of tools available to analysts and marketing departments have become more accessible.

Day after day, critical business rules are being implemented based largely on the identified use of keywords and phrases in free-form text fields provided by prospects, partners and customers, as well as the frequency and proximity (clustering) of their use in relation to each other. It's an analyst's dream, and sometimes a marketer's nightmare: There may be too much data to deal with.

I've often said that text mining is a first cousin to the more established data mining. Both seek to discern patterns and trends from substantial data repositories. Yet it's not easy devising computer systems that can "read" text that is written as natural language. Thankfully, a discipline known as natural language processing is providing success stories.

Marketers are using text mining for two primary functions: classify data subjects by segment and predict behavior. Let's explore the latter use via a personal example.

Customer Service Analysis: A Missed Opportunity
During the past year or so, I have been experiencing problems with my cable service. Initially purchasing the television component of its offerings, I quickly enrolled in the company's Internet service and, not too long after, its phone service. My frequent calls to customer service personnel resulted in the same response with which many readers may be familiar: "Disconnect your cable modem, wait a few minutes and then reconnect." This, I was told, would re-initialize the cable signal entering my home and get me back online.

While I wasn't quite sure what "re-initialize" meant, it did appear that this procedure worked. The only problem was that I was disconnecting and reconnecting several times a day—hardly a quality experience.

With the special package pricing that I was receiving, and the lack of true competition, I had little choice but to grin and bear it. With the entrance of a new competitor, however, my thinking quickly changed. Here is the chain of events that occurred during a five-week period:

• I demanded that a technical consultant visit my home to diagnose and repair the problem.
• I was told that the firm would be happy to send out such an individual, but there would be a charge. I refused to pay.
• My special pricing package expired; it was valid for one year.
• I, in no uncertain terms, made it clear that my days as a customer were about to end.
• I called and asked for an extension to the special pricing. After all, I experienced so many problems that an increase in price seemed ludicrous.
• I was told, "Sorry, we can't do that."
• I told the company, "OK, I will be leaving your firm."
• I called up a competitor and arranged for connection to a new service.
• The new service was connected promptly.
• I spoke to my original cable company to disconnect service.

Please make no mistake. My cable bill was about $150 a month, and I always paid on time. With all the back-and-forth conversations outlined above, it would appear that someone I spoke with should have gotten the message that I was not a happy camper. There were clues in my voice, both in terms of how I spoke, as well as in what I said. So why was the cable firm so unresponsive?

While many firms capture and analyze structured type data, such as transactions, few capture these critical conversations—and those who do never bother to analyze it any meaningful way. In 2010, no business has an excuse for letting such everyday interactions go unanalyzed. By combining these customer service logs with transactions data, this cable provider could easily have determined my intent to attrite, taken a more proactive approach and, indeed, saved a valuable customer. Just multiply $150 by the thousands who may be leaving the cable provider and you are talking about substantial sums.

Analyzing Free-Text Data
Nevertheless, there is a select group of marketers that has been employing these technologies. Presently, text mining is benefiting from applications triggered by the reams of text data accessible to businesses. Well in excess of 70 percent of an organization's data is contained in free-text form.

While not all investigators agree on all the attributes characterizing text mining, four of the features fairly well accepted include:

• Cleaning, which organizes material for subsequent analysis. Typically a major ingredient of text mining, cleaning can take significant amounts of time. Software vendors have made substantial progress in facilitating data text cleansing. Text data usually presents itself in unusual formats. Just think for a moment about the issues that might arise as one attempts to dissect notes from a customer service agent or a variety of Web pages.

• In categorization, text is mechanically classified to one or more categories. If the categories are predefined, then supervised learning algorithms can be used to learn the nuances in the text that distinguish the predefined groups. If the categories are unknown at the outset, then a clustering procedure can be used to classify the text into groups.

• Extraction includes "extracting" substance of interest from the text. Keyword and phrase extraction involves choosing the pertinent words and phrases that identify the nature of the text. The more complex form of extractions involves a process that incorporates parsing the sentence into subject, verb and object relationships. No simple matter for any automated procedure.

• Modeling is probably the most practical of the functions. It marries essential components of the text with more structured data attributes that are typically used in data mining. The purpose of this union is to further optimize prediction or provide for better description.

The Many Applications of Text Mining
While I have highlighted an example of how the cable/communications sector can leverage text mining, numerous opportunities exist for profiting from its value. Market researchers can finally analyze responses to open-ended questions. Product managers may be interested in mining blogs and social media sites to learn about stakeholder opinions and to identify communities of interest, while spotting those individuals most likely to be "influencers" in their own customer base. Contact centers can better cross-sell offerings depending upon the actual words in a customer's conversation or communication. The point is that text mining provides an additional weapon in the marketer's toolbox.

Let me conclude with the epilogue to my cable provider episode. After my cancellation of all services, I received no less than five communications, including a personal visit. I explained to the agent the specific reasons for my leaving. "Had we known," he asserted, "We would have come to your home to repair the line." "Had you known," I exclaimed. "You had all the information you needed to know."

His firm just didn't read my lips.

Sam Koslowsky is vice president of modeling solutions for Harte-Hanks, a worldwide, direct and targeted marketing company that provides direct marketing services and shopper advertising opportunities to consumer and B-to-B marketers. Contact Koslowsky at (212) 520-3259 or via e-mail at sam_koslowsky@harte-hanks.com.

• MORE ON
• DATABASE & CRM