In 2012, Amit Singhal, a senior executive at Google, predicted that within five years, his company would develop a conversational Star Trek–like computer, complete with natural language processing (NLP) and artificial intelligence. In the heat of competition, several other companies made similar bold predictions.

At the halfway point of that five-year period, natural language speech interfaces have come a long way. Advances in the technology have changed how we use our smartphones (Siri), our Web browsers and search engines (Google), and our cars (Ford Sync). Doctors are even using natural language processing to help record and transcribe information related to patients’ visits or medicines. 

And today, “natural language understanding is core to intelligent assistance,” says Dan Miller, founder and senior analyst at Opus Research, noting that “providing consistent, relevant, and personalized responses to customer queries through IVRs, chat, text, and social networks is a crucial part of large companies’ self-service strategies.”

Miller expects to see more intelligent virtual assistants endowed with natural language capabilities. “Nuance’s Nina, Apple’s Siri, Microsoft’s Cortana, and IBM’s Watson are the most conspicuous, but there will be thousands of branded intelligent virtual assistants that can carry on conversations with individuals, making them vital aids for everyday activity and simplifying [consumers’] digital lives,” he says.

Another use case cited by Miller is in the area of language translation, with semantic understanding engines, such as Linguasys and others like it, able to recognize topics and promote understanding in multiple languages. “Perhaps the most conspicuous use of this function was demonstrated by Microsoft recently as it did near-real-time translation of words spoken in one language into another in the course of a teleconference,” he says.

But despite great progress, natural language processing’s potential has not yet been fully realized, and the technology still struggles to climb the peak of inflated expectations. The Star Trek ideal is still something from science fiction rather than fact. 

The good news, though, is that the challenges are no longer technical.

Speech Isn’t the Problem

“Speech recognition accuracy has gotten very good,” says Ron Kaplan, a vice president at Nuance Communications and a distinguished scientist in Nuance’s Lab for Natural Language Understanding and Artificial Intelligence. “It’s no longer an obstacle.”

Today users aren’t surprised when the speech recognition does what it’s supposed to, but rather, when it doesn’t, Kaplan states. “People expect the speech part to work, and it usually does,” he says.

“Natural language understanding continues to benefit from improved speech recognition,” Miller adds. “There is a symbiotic relationship between accurate rendering of spoken words and high-quality speech recognition for command and control, Q&A, and customer self-service.”

Walt Tetschner, an independent speech industry analyst and consultant and editor of ASRNews, agrees. “I’ve had great successes [with natural language]. It can be done effectively. It works well,” he says. 

Natural language, Tetschner says, “offers an interface that is identical to speaking with a human. If the caller thinks that they are talking with a human, then it is natural language.”

Perhaps natural language processing’s greatest potential, though, lies in the contact center. FedEx is already using it in its call center to help customers with everyday tasks, such as scheduling pickups, tracking packages, finding the nearest FedEx location, getting rates, and ordering supplies. Turkcell, a wireless services provider in Turkey, is using natural language processing to handle 100 million calls a year from customers looking to complete a variety of transactions, from account inquiries to service changes.

These deployments, though, are few and far between. Most contact centers still are not using natural language processing. 

Last year, Software Advice, a Gartner company that assists organizations in finding the products that best fit their needs, called the IVR systems of 50 Fortune 500 companies with business models that focus on customer service. Only two prompted the caller for an open-ended natural language response instead of offering a menu.

Tetschner notes that across vertical market segments, speech is included in only 23.1 percent of automated call steering applications. The vast majority of those with a speech interface use directed dialogue, not natural language. 

Out of 1,202 corporate auto-attendants tested by Tetschner, fewer than 100 (8 percent) had a speech interface. Of those with speech capabilities, only four had natural language interfaces.  

“When it comes to true natural language, there just aren’t many implementations out there,” Tetschner says.

“Many call centers have been slow to adopt the technology beyond trials,” says Bill Meisel, president of TMA Associates and executive director of the Applied Voice Input/Output Society (AVIOS). 

In reality, the majority of speech applications in use today are of the directed-dialogue variety. When combined with well-designed prompts and call flows, the higher accuracy of directed dialogue applications can offer a superior user experience, many industry experts conclude. 

Callers can surprise any speech-enabled system by saying things that aren’t on its preprogrammed list of expected responses. Natural language technologies, though, can be more forgiving and understand more caller responses, but they come at a price. Developers of such systems need to be able to predict everything callers might say and build grammars in advance that contain each word and phrase to be recognized. 

This differs greatly from directed-dialogue systems. Because callers are being directed to say very specific options, there are fewer likely responses to account for in the grammar. This makes directed-dialogue systems easier to develop, and they require less time to troubleshoot and test. 

But here, too, technology has solved many of those issues. “The fear was that [with natural language] you would have to build huge dictionaries,” Kaplan says. “Technology has largely overcome this. It’s not so much of a fear-inducing thing where you have to worry that it will be so much work.”

Most businesses will only have to take the words specific to their industries and incorporate them into the dictionary, he says. Kaplan, nonetheless, recommends that companies looking to deploy a natural language interface “be clear about the types of things you want people to be able to do. Know what services you want to provide and how people might ask for it,” he says.

There are many other reasons for not going with natural language systems. Cost is perhaps the greatest. According to Meisel, natural language systems can be expensive, especially if changes need to be made. Changes, he says, usually require more data gathering, testing, and the involvement of vendor professional support. 

“The prices are enormous,” Tetschner says, noting that a natural language interface can cost $500,000 to implement because of all the customization and data that is required.

But, he says, the added costs can be offset over time by increases in customer self-service. Additionally, “a lot of the hang-ups will disappear because customers will stay on the call,” Tetschner points out.

Customer Confusion

Another hindrance is customer perception of natural language technology. “Callers are still looking to talk to a live agent,” Tetschner says. “Automation of any kind is a turnoff for a lot of people, and that is universal.”

Also, when met with an open-ended prompt, many consumers still get confused. “There are plenty of consumers who are not familiar with these systems and need some guidance,” notes Bruce Pollock, vice president of strategic growth and planning at West Interactive, a provider of contact center solutions. 

At the other extreme, some callers faced with natural language prompts tend to say too much, which also challenges the system. “If the nature of the call is more urgent, people tend to go on at great length about why they’re calling and explaining their problem in greater detail than is needed,” Pollock states.

For these reasons, some companies that have field-tested natural language IVRs have backed them up with directed-dialogue interfaces just in case the natural language engine fails or the caller has other problems. An example of this is the following: “I’m sorry. I did not understand you. You can say, ‘check my balance,’ ‘make a payment,’ or ‘update my account information.’”

While those types of backups are useful to customers, they have done little to advance the cause of natural language processing and have left companies wondering why they should invest in natural language if they still must rely on directed dialogue.

NLP Is Not for Everyone

Another factor that has stood in the way of more widespread adoption of natural language is that not all IVRs are suited for it. “[Natural language] is not necessary if you’re going to present customers with a small list of menu options,” Julio Murillo, a speech technologist at West Interactive, says. “You might not need a natural language interface. You probably just need to reorganize the prompts and shorten the menu options.”

In some cases, a graphical interface might be more effective. “If a caller wants to find a movie, it doesn’t make sense to have a system read off information about fifty different movies that are playing at the time,” Kaplan contends. “That can be presented visually.” 

Where natural language is most effective is with companies that have long lists of products and large numbers of reasons for customers to call them.

That was the case at TalkTalk Group, a United Kingdom provider of Internet, TV, and phone services. Last year, prior to implementing a natural language call steering IVR solution from Nuance, the company reviewed more than 30,000 calls and identified more than 300 call types. 

So far, its call recognition accuracy has been 94 percent, resulting in 16 percent fewer transfers and a 26-second reduction in the amount of time spent in the IVR. Self-service use has increased by 28 percent and customer satisfaction has increased by 0.6 percent. TalkTalk expects to reduce costs by roughly $5 million per year as a result of the implementation.

With those kinds of numbers in its favor, industry insiders expect adoption of natural language to pick up. “I suspect that as the technology becomes democratized, cheaper to implement, and more effective, it will be more prevalent in contact centers,” says Jonathan Gale, CEO of NewVoiceMedia, a provider of cloud-based contact center technology.

“More rapid deployment is on the horizon,” Kaplan adds. “In the next year or so, you’re going to see a lot more of these [deployments]. It’s definitely coming.”

Pollock expects the technology to advance beyond just the opening prompt as well. “Once the customer initially identifies where he wants to go, companies might use natural language for the next level of support too,” he says.

Beyond that, there is a wide variety of research happening around natural language in the labs at companies such as Nuance, IBM, Apple, Microsoft, and Google, as well as in the academic community.

 Natural language, Miller points out, “will only get better.” 

“By design,” he says, “natural language understanding is closely mated to machine learning resources, big data storage, and deep analytics.”

That is precisely the focus of much of the research that Nuance is doing in its Lab for Natural Language Understanding and Artificial Intelligence, according to Kaplan. “Natural language is growing and migrating beyond just voice technologies,” he says. “There’s a blending of natural language and artificial intelligence so that [systems] can interpret customer intentions and act on them.”

To that end, a lot of work is being done around contextual understanding. “There can be ambiguity in interpreting an utterance,” Kaplan states. “It relies on understanding the context, being able to match the concepts with the information available so the system comes to the right conclusions.”

That, he adds, needs more information than the language itself.

“The evolution that is going on now is to have mixed-initiative dialogues,” Kaplan says, noting that these interactions involve a series of collaborative questions and answers between the customer and the system. “The system should be able to reason what the customer wants and draw out other information to help him,” Kaplan says.

Artificial intelligence will also be able to help guide systems when the original customer request cannot be fulfilled. This could be useful, for example, when a customer calls a restaurant to make a reservation and no tables are available at the desired time. “The system should be able to offer other suggestions around the original intent once it recognizes that the request couldn’t be fulfilled,” Kaplan says.

As another example, Kaplan says customers should be able to ask for a romantic dinner and a movie on a given night and have the system know that a romantic dinner probably involves a white tablecloth and candles, that dinner probably takes about an hour and a half, what movies playing might be considered romantic, and then recommend a theater close to the restaurant. “Some of these capabilities are on the horizon. Many of the core technologies are available; it’s just a question of how to put them all together,” he says. “We’re on the cusp, with a lot of new capabilities coming to make natural language better for users,” he says. “It’s in the lab, and it will come into products in a year or so.”

It might not be Star Trek yet, but the industry is getting closer. 

News Editor Leonard Klie can be reached at lklie@infotoday.com.