NLP in Your Contact Center: What You Need to KnowNLP in Your Contact Center: What You Need to Know

As I mentioned in my previous post, "AI in Your Contact Center: Understanding the Basics," the use of natural language processing (NLP) to conduct a conversation with a customer is far and away the most impactful use of artificial intelligence in a contact center. NLP underpins the conversational AI used with Web chatbots, voice assistants for enhanced IVR functionality, and voice computing interfaces for virtual assistants like Alexa, and as such is a critical technology to understand for the modern contact center.

NLP comprises two component parts: natural language understanding (NLU) and natural language generation (NLG). NLU involves mapping of natural language input into useful representations for processing and analyzing, while NLG turns representations into natural language.

In some respect these technologies are old. Computer scientists Robert Mercer and Peter Brown filed their first NLP-related patents in 1972. But with the availability of affordable, high-powered computing platforms, these technologies can now support real-time responses fast enough to achieve comfortable human interaction and be usable in customer engagement. In practical terms, this means the ability to deliver conversational responses in less than 300 milliseconds.

If conversational interfaces are your goal, then you should use NLU and NLG together. However, you can still find utility in using these components separately. Asking an IVR to transfer funds from one account to another only requires NLU, for example. The audio confirmation is what requires NLG.

Understanding NLU
The following are descriptions of some capabilities available with NLU platforms.

   Discern Intent: The intent of a customer can be discerned by the processing of a complete utterance. However, grammatically complete sentences aren't necessarily common in the customer engagement environment, so you'd need to use contextual and metadata elements to manage the conversation to a successful conclusion. When trying to retrieve a stock quote, for example, a customer may ask, "Tell me what Alphabet is trading at right now." The machine captures this information, applies contextual understanding, and responds, "Alphabet is trading at $1055.82." It's discerned the customer's intent to retrieve a stock quote.

   Disambiguation: Customer engagements are less prone to uncertainty than some other forms of conversational AI, and so tend not to require disambiguation as often. This is because interactions focus on a group of customer attributes, products, or services and not the universe. This doesn't mean a customer will never say something ambiguous or contradictory, like "Ship my order to my home address, no, on second thought, please send it to my work address." This leaves the machine to disambiguate the confused messaging. In these types of cases, you would want your AI assistant to recite a confirmation: "Just to confirm, you want your order shipped to your work address?" This confirmation step will provide the machine with an affirmation of correctness it can use to prevent further misunderstanding and aid in the machine learning process.

   Terminology Extraction: Every industry has some unique terminology. The tech industry is full of it, an example is "port" in telephony. If a customer wants to move a phone number from one cellphone to another, he might say something like, "Please port number 410-555-1212 from T-Mo to the following SIM card 123-4545-4545-9865." With the proper understanding, the machine will extract the word "port" and the SIM value from this sentence so it can execute a robotic process that will move the number to the new SIM card. Similarly, the machine will know that "T-Mo" refers to the carrier T-Mobile, and will initiate contact so the number gets released to the new carrier.

   Translation: Trying to use translation in the customer engagement environment typically isn't a good idea. Rather, the best option is to select an NLU/NLG solution that handles languages natively. Translating a customer request from German to English, then processing a response in English and translating it back to German, for example, is a formula for aggravation.

   Parsing: A popular example of NLP parsing is in the interpretation of the sentence: "I saw a girl with a telescope." Did I use a telescope to see the girl, or was the girl I saw holding a telescope? Customer engagements rarely require parsing sentences to determine meaning, so coming up with an example is difficult. As such, a parsing algorithm is more of a like-to-have than a must-have NLP feature for customer engagement platforms.

Continue to next page

Continued from previous page

   Stemming: This process comes into play for interpreting miswritten or mispronounced words, as well as for reducing time needed to program a machine to understand intent. Consider the following: "I want to have funds transferred." The stem of "transferred" is "transfer." If you tune the machine to consider all forms of "transfer," then you can save time by not having to program every form of the word manually.

   Named Entity Extraction: I touched on this earlier in my stock trading example for understanding intent. After the customer asks about the stock price for Alphabet, he might add, "What about Amazon?" The machine maintains the context and says: "Amazon is trading at $1522.32." Taken on its own, "What about Amazon?" might not refer to stock price, but rather be a request for information about the rain forest. Using named entity extraction, the machine is able to deliver a response that matches the context of the conversation.

Some NLP solutions come with detailed lists of publicly used named entities, such as company name as in my example, but others don't. Even with a provided list, the named entities in your business may be different from others so you'll need to pay special attention to building your list. Besides company names, the named entities list should include the names of products, cities, countries, vendors, and processes -- include any named entities that might come up and would help the machine deliver a faster, more accurate response during a customer interaction.

   Topic Segmentation: Building a knowledge base traditionally has involved a manual process for curating content and determining topics and subtopics. NLP solutions can automate this work with a topic segmentation process that determines which sections of a document apply to a specific customer request. For the purposes of speed, it's best to execute topic segmentation in advance and generate tags on all knowledge content so that your machine can more quickly present the correct knowledge when customers request it.

   Discern Sentiment (Emotion): Sentiment analysis has many uses in a contact center. As the linchpin in an offer management solution, it's hard to duplicate. If the customer is happy, make him an offer; if he's not happy, then conference in a supervisor to get the issue resolved. Analyzing words and punctuation can help determine sentiment in a text interaction. Voice interaction add pitch and volume for use in sentiment analysis, while video interfaces bring facial expression into the analysis. Some vendors offer all of the above, but many do not. If you're interested in sentiment analysis, then make sure you understand your selected vendor's abilities in this area.

   Summarize Content: Summarization is an interesting tool in the customer engagement environment. For its Watson NLU engine, IBM charges $0.003 to summarize a 10,000-character document and the same price to summarize an eight-word utterance from a customer. NLP solutions aimed at customer interactions are typically tuned to these shorter utterances.

In customer service environments, summarization is more of a statistical tool. It can effectively take the place of call disposition processes, providing contact center operations managers near-real-time insight into why customers are calling. I know I would have found an NLU summarization tool useful back when I was running a pharmacy benefits operation. For example, one particular day at 4 p.m. the phones started ringing at triple the historic rate. It turns out that this is about the time retired people go to their mailboxes, and one of our brilliant marketing people had sent a mass mailing with the first line reading: "Your benefits may be in jeopardy." Once we figured this out -- two hours later -- we quickly pushed out a script so all agents could handle these calls more efficiently. If we'd had an NLU summarization tool, we could have identified the trend in minutes.

   Tag Content: Tagging content is necessary when a contact center uses large volumes of unstructured data to support customer interactions. Searching on tags is much faster than having to search all content in a knowledge base, which means agents -- virtual or human -- can deliver the correct answers to customers more quickly than they can without content tagging.

   Taxonomy (Classification): Classification is used to identify certain types of words. If you're in the mortgage servicing business, then Columbus probably refers to a place, not a person. Similar to named entities in purpose and functionality, NLP-based taxonomy can be useful for enterprises that have very broad product catalogs. These tools can reduce the time and confusion that a customer may engage when dealing with thousands of products.

Getting a Grip on NLG
In customer engagement, NLG is typically understood as an elaborate directed-dialog implementation. The responses or follow-on questions the interface exposes to the user are created using NLG. Specifically, an NLG processor exposes text to the user (as in Web chat) or to an intermediary technology like text-to-speech (as in IVR or voice computing) that generates speech the user can listen to.

If you intend to use a voice interface, then you should put some thought into the quality of the text-to-speech solution. Punctuation, gender, energy, stress, phoneme length, intonation, syllabification, and tone can all contribute to the quality of concatenation you expose to your customers. While you may not necessarily want to explore the algorithms behind each of these factors, you'll definitely want to listen to the interface and determine what sounds good to you and your team.

Options Galore
If you're still reading, then you're probably considering an NLP implementation in your contact center. You have a lot of options, with more than 1,000 companies now offering NLP services. Some have pre-built functions that support the features I described above, while others require you to build your own features. The good news is that as contact center technologies go, none are very expensive to use. However, some can be more expensive to implement and manage than others. Selecting the platform or product that best supports your business use cases is the key to success.