A Reality Check On Natural Language For Conversational Computing
Conversational computing has our attention. We have been talking to machines since as far back as 1961, when the world’s first digital speech recognition system, the IBM Shoebox, was released. The Shoebox could recognize digits and a grand total of 16 words. Cut to half a century later: Humanity is trying hard to fall in love with an entire class of loquacious devices and systems. These speech-based systems are often novel, sometimes interesting, but generally of marginal utility. They can perform limited tasks and are largely good for narrow, directed use cases.
Enterprises are hard at work trying to replicate these speech-driven interactions in their contact centers and enterprise systems (through NL-IVRs, voice assistants, and chatbots). Yet the killer app for natural language in the enterprise is yet to emerge. The road to enterprise natural language is littered with the bones of dead chatbots and underachieving NL-IVR investments. Consumers have mostly driven past, underwhelmed.
A few thoughts:
- Understanding is the weakest link. First, let us get the terminology right. Conversational computing is the domain of natural language processing (NLP), an overarching umbrella that covers the recognition, understanding, generation, and synthesis of human speech. Of these, the first (speech recognition) is fairly evolved across a wide swathe of languages. However, natural language understanding (NLU) is the weak link. For enterprise speech to be useful — to make it a truly useful part of the contact center interaction mix — it needs to speak in more humanlike ways and solve real consumer problems that are usually more involved than looking up the weather or turning the lights off. Rule-based hacks can only carry us so far.
- Yes, we have no linguists. Computational linguistics is a critical skill set required to improve the quality of NLU. Yet this is a scarce talent, and this scarcity puts a hard limit on the amount of focus that is put on understanding and decoding language (and, indeed, languages) at a semantic level, which is useful to computers. Indeed, consumers in various global markets speak in different languages, and the gap in maturity of NLP for different languages varies significantly. And this gap isn’t really shrinking.
- Keep calm, and set pragmatic expectations. Here’s a harsh reality: Conversational AI may be getting better, but automated customer conversations are not. Most clients do not know what to expect from natural language investments. Instead, promises by vendors of delivering speech nirvana have delivered largely lukewarm results in the real world. In this context, how do you build a business case that captures the coverage and value of your investment in conversational computing and NLP? How do you build a road map for delivering useful conversational customer experience in the markets that you serve?
My latest report, “The Six Factors That Separate Hype From Hope In Your Conversational AI Journey” (Forrester clients can access it here), has pragmatic advice for enterprises looking to embark on a strategy to deliver natural language interactions to their customers. I would be delighted to hear your feedback or to listen to your experiences and stories.