This post was originaly published in Chatbots Magazine.
Language technology is all the rage across the world of innovation these days. Conversational interfaces, machine translation, natural lanuage understanding, chatbots — the buzzwords abound. The hype is not undeserved either. Truly interactive, powerful conversational interfaces could fundamentally change our relationship with technology by intermediating our every interaction with technology in all but the most specific use cases.
The first step down the road of “conversational intermediation” is the rise of chatbots. Chatbots have quickly gone from musings and cool-but-impractical demonstrations to widespread in various forms. Their rise comes down to several interlinked causes.
The Rise of Chatbots
With the growing penetration of SaaS applications into every level of business and the deepening pockets of entrepreneurship globally, several factors have combined to push chatbots into the vanguard of the language technology revolution sweeping the world.
The first of these factors is the rise of cheap, easily available, and powerful computational resources powered by platforms like AWS (Amazon Web Services), Microsoft Azure, and others. Neural networks, complex algorithms and high-volume data transactions have become vastly more scalable and widespread as a result, enabling acceleration of research and broad experimentation with chatbots. The deployment of graphical processing units (GPUs) has also been an essential enabler.
The rise of chatbots also owes a debt to substantial progress in language technology, particularly with the advent of deep learning and neural networks. Language technology dates all the way back to the dawn of computing with IBM and Georgetown University’s successful demonstration of a machine translation system in January of 1954. However, progress in the field was limited and it wasn’t until Google applied its web-crawling muscle to the problem that progress started to accelerate. Google crawled billions of webpages to find websites with versions in multiple languages. This vast dataset of translations helped train the translation models underpinning Google Translate. Indeed, deep learning and neural networks has allowed companies to turn vast data sets into learnings, and learnings into intelligent systems.
A third major cause of the rise of chatbots has been the spread of chat-based messaging platforms like Slack, Hipchat and even Facebook’s messenger. These communication platforms play host to chatbots and present a series of natural use cases for chatbots to solve. However, it would be difficult to imagine chatbots spreading widely if email continued to dominate all professional and private communication. Notwithstanding bots specifically deployed on email platforms, email as a medium is detrimental to conversational interfaces. A conversation with long delays between responses isn’t really a conversation.
Finally, the rise of chatbots has been supported by the open-sourcing of bot frameworks from the leading tech giants. Microsoft, Amazon, Google and IBM all offer chatbots and/or their core language technology to the public to experiment with and build off of. This counteracts one of the primary drawbacks of deep learning — vast datasets are needed to train models and neural networks. These openly available frameworks enable entrepreneurs and businesses to directly experiment with chatbots, without the vast uphill battle of getting enough data to train their NLP/NLU models. Plenty of open source Python libraries (among others) also exist that support the NLP behind chatbots.
Bots in the Field
While we’re still in the earliest days of chatbots, there are three categories of bots that are visible, active and of day-to-day value as it stands: enterprise, productivity, and knowledge-focused.
“Enterprise bots” are deployed by corporations for specific, usually customer-facing, use cases. The best example of this is KLM’s Facebook messenger chat bot. Users can install this on their Facebook messenger app and receive push notifications with details on their flights, check-in times and more. They can also interact with the chatbot conversationally to pick seats on the flight among other functionalities.
“Productivity bots” aim to automate routine tasks like scheduling meetings or finding a place to go for lunch. They are can be deployed in either team-based messaging platforms like Slack, or in personal communication platforms like email. Meekan is an example of the former, as it focuses on connecting different team members’ calendars together. On the other hand, X.ai is an example of the latter, and focuses on managing an individual’s personal calendar with minimal conversational interaction.
The final category of bots is “knowledge bots.” These are bots that often come with more robust AI capabilities and are designed to build knowledge graphs and make them queryable via question-and-answer. These knowledge graphs can be very deep or relatively shallow. IBM Watson is an example of a very deep knowledge graph with a conversational interface over-layed.n the other hand, Rover offers a much lighter knowledge graph of your organization in Slack. Although Watson and Rover are on opposite ends of the spectrum, there are also bots that fall in between. For example, Visabot, which helps US immigrants by providing details on various visa programs via Facebook messenger, is another example that falls somewhere in between.
It’s no secret that natural language processing has made stunning leaps in the last few years and no longer presents a serious challenge to creating intelligence in software. However, to build truly intelligent, conversational chat bots with personalities we will need significant progress in the areas of natural language understanding and natural language generation. These two areas are significantly more complex problems, and progress is not guaranteed. In a recent Technology Report on Language Technology the Economist aptly described the challenge that “machines cannot conduct proper conversations with humans because they do not understand the world.”
Natural language understanding (NLU) is a hard nut to crack. However, we’re likely to see progress when NLU is required in specific use cases. For example, question-and-answer bots will become ubiquitous in customer service, customer support and FAQ roles both internally and externally for businesses. In this case, the range of semantic understanding needed to produce reasonable NLU is limited and domain-specific and thus easier to train on. It’s a completely different challenge for us to ask a question-and-answer system what someone else thinks about a question without having that data already on hand.
In NLU the many layers of meaning that make up human life and society presents a massive challenge for conversational interfaces. It may even be an insurmountable challenge, chatbots may simply remain limited to specific use cases, and we may not see a truly generalized conversational interface — just like we may not see generalized AI.
While many NLU & NLG challenges remain ahead for conversational interfaces, it’s clear that they are going to become very common because of their cost-savings and consistency in service. From the perspective of venture investors, we’re still looking at this trend, seeking to understand possible business models and put it into context in the wider evolution of the technology industry. It’s early but exciting days for conversational interfaces.