The Language Technology Industry Summit 2018 (28-29 May in Brussels) will illustrate how language technology addresses challenges in artificial intelligence. AI leader John Rauscher, outgoing CEO of Yseop, will provide insights and case studies from the NLG (natural language generation) sector. Philippe Wacker has spoken to him in advance of the Summit.
Philippe Wacker, Secretary General, LT-Innovate: What were your main professional career stages and achievements, especially in the language processing area? What was it that interested you about this domain?
John Rauscher: I was lucky to launch three start-ups from scratch in my career. I led the first one to an IPO, sold the second one to Oracle, and bootstrapped the third one from the invention stage to one of the top positions in their market. The latter is still growing and focuses on Natural Language Generation (NLG). I have been fascinated by what NLG offers for three reasons: I am the author of two published business books and I know how much time it takes to write great content that everybody understands. With NLG, you can write 2,000 pages of commentaries per second. The second reason I am passionate about NLG is that, a long time ago, I wrote an essay on the business impact of Gutenberg’s invention. My conclusions were that France and Italy own him the shift from monasteries writing books as a business model to monks producing wine and cheese! Finally, I believe that language is the #1 issue that prevents human beings from delegating repetitive service tasks to collaborative robots.
Philippe Wacker: What were the main challenges you had to overcome when designing language systems: business, hierarchy, technical, cost, personnel, etc…
John Rauscher: The main challenge was to start from an invention and find out the best use cases in order to transform the technology into an innovation. As often in the early market of the “next big things,” engineers create a concept and then ask themselves what the best application is. It is impossible to ask potential customers what they think about it early on as they could not imagine that this capability would even exist one day… Of course, in an ideal world, we should start by identifying the pain before writing the solution.
Philippe Wacker: References are regularly made to NLG vendors’ websites (Automated Insights, Arria, AX Semantics, Narrative Science, Narrativa, and Yseop). What differentiates the products of these vendors from each other? How do you see this market evolving?
John Rauscher: They are actually quite different from each other and nobody covers the big picture. Arria NLG and Yseop are the two well-known European players. Both carry strong NLG technologies. Yseop was the first to generate multilingual content and to offer a self-service interface. Based on what analysts wrote, Automated Insights is more of a template-based approach than real NLG per se. But who cares? It depends on what you need to achieve: For generating short content for the web, it may be enough. Narrative Science is the best-known player in the US market because of its contracts with the US Government. One of its investors is In-Q-Tel, the CIA’s financial arm… No doubt that they are strong where customers want highly secure NLG applications. Narrativa focuses on the Middle East market and I have never met them in a competitive environment.
Philippe Wacker: What arguments do you use to explain the use cases for language processing apps to potential clients? Are these changing as time goes on and as digital transformation is more common?
John Rauscher: As your question implies, sales points vary depending on the target use case. Generally, the quality of the content is the #1 selling criteria for using a NLG solution. Of course, for a NLG solution to be buy-worthy, the final customer needs a large volume of reports to generate. A requirement for a large volume of personalized reports makes the ROI easy to demonstrate. The third sales point is the ability to create new business models such as personalized Web pages depending on each Internet user, articles comparing different products that are generated instantly, diagnoses with the appropriate level of explanations depending on the context, etc.
Philippe Wacker: Our language area - cognitive/linguistic/speech/ambient, translation, business process optimisation etc. - is very hard to decipher for many people. Do you think there are ways of improving communication about language processing tech in the future?
John Rauscher: Yes, I do. I always start by explaining the four different types of information human beings are able to understand and memorize: (1) raw data such as the PIN number of your credit card, your name or any number in an Excel spreadsheet for example; (2) information in natural language – this is what you mainly search using Google, i.e. any text with a meaning; (3) knowledge that has been in books since Gutenberg’s invention - knowledge is how to do things whatever the topic; and (4) expertise that is actually knowledge applied to a specific context, for example a question from a customer.
Explaining these four types of text allows to simplify the understanding of the terms described in your question. For example, generating “expertise” automatically requires applying reasoning (cognitive) based on business rules (knowledge or business process in your question) to data. Also, the challenges of translation depend on the type of text.
Philippe Wacker: How do you see the impact of Machine Learning on the field you have been involved with and do you think that AI/ML will transform our ideas about language applications and how to build them?
John Rauscher: As far as I know, nobody has been able to launch an NLG platform only based on Machine Learning/Neural Networks. ML as of today implies a statistical approach. To generate language, one needs a deterministic approach because any language is based on predefined grammar rules. You cannot generate text based of the average quality of text generated by human beings… The average quality of text is not what customers are looking for… I believe that the future base technology for NLG will be XAI, i.e. explainable machine learning.
Philippe Wacker: Debunking the myth of AI/ML as a job-killer, you stated in a previous interview that “technology has the potential of turning any knowledge worker into an expert”. Could you elaborate on this?
John Rauscher: Over the last 20 years, our daily work has changed. We specifically do tasks in three different areas we did not do before: (1) Knowledge workers are required to analyse large volumes of data, way above what their brain can process. Look at a CRM. You generally have access to hundreds of data points. Experts in cognitive science tell us that human beings cannot analyse more than 30 data points at best during a decision-making process. (2) The portfolio of products and services that companies market, sell or support has multiplied tenfold in 15 years. There is too much to know for each knowledge worker to do a good job anymore. (3) We cope with more and more regulations that complexify our ability to advise a customer properly. I believe that AI will allow every knowledge worker with no specific experience to (almost) become an expert overnight. The three tasks described above will be done by AI and human beings will focus on customer relationship, empathy, and creativity, not on trying to analyse large volumes of data, comply with business best practices and regulations. AI will augment people with less stress and they will be happier at work. Creative people will be paid more than hard workers.
Philippe Wacker: You also said about innovation “if there is no wave, there is no chance to make it happen”. Where do you see the next wave of innovation coming from in our sector?
John Rauscher: I believe that the next wave will be automating an intelligent dialogue. So far, what we see with chatbots is bi-directional monolog, not dialog. I have never seen Google asking me questions to finetune my search! From what I have seen, when chatbots ask questions, they do it through pre-scripted questions built in a decision tree. This is far from what customers want. A successful dialog requires at least three components: reasoning, NLU and NLG. Many technical teams around the world work on this challenge but there is still a lot to be done. It would be a huge step to see a good dialog solution working seamlessly.
Philippe Wacker: You have just stepped down as CEO of YSEOP. What are your personal plans for the future?
John Rauscher: There are many great disruptive solutions coming out in the language space every month. But often, the missing ingredients to make things happen are marketing, sales, international, and management of best practices. Money is needed but is not enough to turn the invention into a great innovation. I am always interested in learning more from disruptive technologies, but I know that the secret recipe for success is the ability of two people, the technical founder and the market founder (where I fit) to work together well in the long run. I look forward to meeting great technical founders in Brussels at the end of May.
