The Speculation on Conversational Ai

Conversation has been an unequivocally human ability, differentiating us from the rest of creation, being the basis for culture and civilization, and defining our unique level of intelligence as a species. It serves many purposes in our everyday life: communication, coordination, social connection, completion of complex tasks, education, comforting, and entertainment – to name a few. Nothing like a good heated dialog to get us all excited and creative about a topic or … each other. A rapid fire repartee makes us laugh or persuades us who to choose for our next president! Humans are really good at making conversation whether they are for work or just shooting the breeze. When we talk to each other, we are constantly leveraging contextual information and knowledge to convey sarcasm, read between the lines, and express our personality. Conversation is a high bandwidth channel where knowledge, instructions, behavior, emotions, strength of will, and many other messages are conveyed via language (written or spoken) and its structure. In our daily lives, we typically engage in spoken conversation, whereas recent technological innovations have introduced to us the habit of almost instantaneous, typed chatting.

The complexity of presentation, the structure, the conduct of the interaction, and the amount of information it carries are all measures of intelligence of the participants. We are surrounded by an animal world of inarticulate cries conveying simple messages, and by a stone cold world of systems and machines that require specialized and limited forms of instruction and manipulation. Thus, it has been humanity’s dream to interact naturally with tools that will do its bidding. Popularized in Sci-Fi novels, the concept exists since Homer’s days (see Ulysses Rhapsody Σ where Vulcan is served by human-like handmaids maid of gold). AI comes thousands of years later with the promise that it will make humanity’s dream come true: have “intelligent” conversations with our machines, in the sense that we will be able to get information, pass instruction, acquire education, or even get advice in a natural way. Actually, the measure of success of these AI conversational systems is none other than the Turing test, which is also a measure of intelligence: to have a conversation with a machine on a topic or task that would be indistinguishable from talking to another human. However, apart from human narcissism and creativity, what fuels further this endeavor for intelligent conversational systems and assistants is a long list of enterprise applications and strong market need for personalized conversation between businesses and their customers. So far, the complexity and limitations that existing dialog tools and resulting conversational systems suffer from, have served business and their customers with disappointing experiences. They get the job done eventually, but after great effort, high cost for development and maintenance, and relatively limited human level interaction experience. This is because today’s automated conversational systems aren't actually intelligent! Designers with domain knowledge and computer expertise define and program each conversation with the scripted responses that users can expect when interacting with the automated dialog systems.

Thus, conversational systems are built based on a decision-tree logic, where the response given by the bot depends on a dialog state defined by specific intents and keywords identified in the user's input. IF user's input contains 'shop' or 'buy' (intent); AND ‘cellphone’ or ‘mobile’ (product type) ;THEN send message with cellphone listTypically, designers have to program 3 major components in order to make an automated conversational system: a) its natural language understanding part, i. e. , the part that parses and analyzes human language and identifies the parts that are important for the task, b) the dialog management part, which basically identifies a state for the dialog based on the history and the current parsed input in order to decide what to do next, and c) a response generation part, where typically the designer programs the scripted responses of the system. What all this means is that the resulting systems will seem as intelligent as the effort (and patience) that was put by the designers who created them: capturing and anticipating a large number of potential use cases and inputs, creating appropriate and natural sounding responses.

Furthermore, adapting and maintaining such rule-based dialog systems with changing information or new information about a task is a labor intensive and time consuming programming job. To mitigate these shortcomings, the next generation dialog systems need to be able to learn. For starters, there are two easily accessible sources of knowledge:a) examples of human-to-human interaction, and b) existing data (books, websites, manuals, databases). This is after all what business also have at hand. Companies have collected huge amounts of example conversations from the interaction between their agents and their customers over the phone or other channels (online, twitter, etc. ). Similarly, companies have abundant organized (databases, knowledge graphs, websites) or raw (documents, manuals) data that contain knowledge related to business operations and a variety of business services (booking trips) or goal-oriented tasks (maintenance, repair). However, several attempts so far for automated example-based systems have not been successful and have resulted in unexpected or even comical results (remember the offensive Microsoft Tay and the amusing Facebook negotiating robots challenge).

The reason is that you have to be careful what data you feed the system when you train it and also what mechanism you use to produce answers.Such systems will be accepted in the business world only when their responses are not free-form, but can be restricted within a set of responses acceptable by the business. What the above highlights is a deeper truth: automated conversational systems today are missing the link with available knowledge. This connection and knowledge transfer from data is provided so far directly by human design and programming. Thus, when building an automated dialog system we need to recreate existing functionality from scratch, i. e. , recreate our website experience in a different way, while the underlying information is the same. That means extra work for businesses to create a totally different channel to handle customer requests – which also results in lack of consistency for the user. By contrast, human agents do not have that problem. They are able to continuously acquire knowledge by observing others or from documents, assimilate new information, and appropriately organize so as to augment their capability to conduct new goal-oriented conversations. Transferring this human ability, even to some extent, to our automated dialog agents and Thus, learn from example dialogs and existing knowledge sources, is necessary to make progress and it clear that the market wants it. If we can appropriately represent knowledge about a goal-oriented task and automatically integrate it in neural network or traditionally programmed dialog systems we expect to have several benefits. First of all we expect to improve performance/accuracy of response of the systems. No more dependence on how thorough and expert the system designer is; the system will learn all there is to know from available data. Ever more so, we will be able to instill some “common sense” to the systems that can carry over from application to application. The systems will be able to adapt quickly to new and changing information as well as operate and evolve in the absence of example dialogs. Still this connection of dialog to knowledge remains elusive – it is a hard problem for machines. Taking advantage of knowledge and restricting conversational systems to the set of acceptable responses will dramatically automate their development and maintenance, but this does not mean that we do not still need to spend effort in programming them or improving their language understanding capabilities. Conversational interfaces represent a big shift in the way we are used to thinking about interaction with our “dumb” computers and “smart” phones!

Conversational computing is a paradigm shift that requires designers to change their thinking, their deliverables and their design process in order to create successful bot experiences. We expect, therefore, that major progress will be also be made in the systems and tools that allow the composition and integration of different components related to dialog. Designers need to be able to take advantage of the strengths and facilities that different AI technologies provide either in analyzing human language or in conducting dialog learnt from different sources. Thus, over the next couple of years we will see a proliferation in the market of tools that will not only support building conversational systems from scratch, but will also allow evolution to similar tasks, carry-over of knowledge and dialog management, and continuous incorporation of new data. Such conversational interfaces will grow to be the new operating system or digital mesh that will hold technologies together. Our future will be flooded with digital assistants, drones, robots, and self-driving cars. Therefore, we also need to look toward innovative ways to converse with these new devices. That means not just giving one-way instructions or queries, but conducting two-way interactions that meet our needs. That’s where conversational computing comes in. We need conversation not only for form filling or step-by-step instruction, we need it because we do not know the ever changing options (e. g. tickets and dates available, or the new situations encountered) and the systems do not know our needs, preferences, or do not have our special training and wisdom at any given time, to complete a task. Big companies are making big investments in the conversational area: Google, Apple, Amazon, Facebook, IBM, Baidu - to name just a few. And by mastering conversation, they can master the world. The next Alexa will be your home assistantor your hotel concierge. The next Siri or Google-Assistantwill be your personal assistant at the office and home. Facebook will interact with you just like one of your friends. But apart from the domination in our personal everyday lives, conversational systems will take over the business world, since they will provide faster, better, cheaper customer service. That’s were companies like IBMand many start-ups are making their play. There are estimates from Gartner that AI will account for 85% of customer relationships by 2020, and recent market analysis indicates that today 60% of regular customers (i. e. , you and me) would prefer to talk to an automated system than a human to complete simple tasks, if it is faster and more informative. However, the majority of us (more than 70%) still do not trust automated systems with complex tasks or with our money. Also most of us, the survey says, do not want to rely on automated assistants that take decisions for us. Those cases still require the human touch, somebody who understands our needs, can negotiate, is able to explain, and lead the conversation to a win-win solution.

Major strides are expected, therefore, in systems that demonstrate more human like characteristics and that take into account more modes of interaction than just typed messages. Microsoft for example is working on a natural user interface (NUI)that combines natural language with gestures, touch, and gazes, to help deepen the system conversations. Everything can be “felt” by sight, touch, or sound. That’s the kind of multimodal conversation that will make automated conversational systems more human-like. Google recently demonstrated Duplex, a concept assistant that makes appointments and reservations for you: it sounds and interacts very human like. AI can play a major role in this. Research is already prototyping deep learning for conversational systems that is increasingly “deeper”: Instead of learning dialog from textual dialog examples, novel AI systems are in the works and will learn directly from spoken interactions. MILA, the research dialog team from the Un. of Montreal, Facebook, Samsung, Microsoft, and Google are already working on that direction,,,,. This is going to be very powerful. Recall from our younger years that spoken dialog was our major and only skill to learn, to play, and to teach others. We were able to negotiate with our parents before we were able to read and write, and we were able to describe how to play a game to our peers or coordinate to play it, before we went to school. Spoken dialog interaction is a rich form of intelligent communication that only becomes more sophisticated and complex over the years: it continuously incorporates what we learn from our interaction with others or from knowledge sources (books, documents, articles, manuals, etc. ) Are we close to seeing such systems any time soon? “Patience you must have, my young Padawan. ” Our current experience is with concatenated systems, i. e. , systems that have a speech-to-text component that transcribes our speech which is then passed to the conversational system and the response is read back to us. Such systems often make mistakes and get confused, resulting in negative customer experiences.

One of the reasons is that they are not tightly coupled: information about what the conversational system expects is not passed back to improve the transcription system, and vice-versa an erroneous transcription is blindly passed to the conversational system. Learning directly from the spoken dialog examples can address many of these problems and also incorporate new aspects in the automated dialog intelligence: emotions, attitudes, expression styles, voice inflections. We will be talking to a machine that is able to understand better certain elements of our human nature and has been programmed or “trained” to respond and evoke appropriate human-like characteristics. Our interaction with such bots will not be strained and suspicious but natural and assuring. If I were to make a bet in the market and upcoming technologies, I would definitely put my money on three kinds of companies: a) those that innovate in the area of AI for automatically creating dialog systems from available data and human-to-human interactions - they are going to be the basis for the next wave of conversational products, b) those that have platforms that allow easy and fast composition of diverse dialog related components and their integration into applications, and c)those that use the above advanced AI algorithms and integration platforms to build evolving vertical applications that enterprises need.

These three types of companies provide the three necessary layers / pillars to the AI conversational assistants market: core technology, development middleware, and application layer. Look out for the big players and for many startups that cover one or more of the layers. They will be the next wave of successful companies in Conversational AI. Should I fear the future? Are AI conversational assistants coming to heartlessly take away the allowance of the high-school kid who takes the phone orders at Domino’s in the summer? Are they going to take away the bank teller’s job or the investment broker’s living? Are they going to take my job? In part, these fears are well-founded. But many of these jobs are already disappearing with the advent of e-commerce, web-banking and investing, etc. It’s happened before with every technological progress in the past, but what makes us more wary now is that this wave of progress it is touching on elements of our human nature and intelligence. However, do not be quick to discount humans yet!

Many more new jobs will sprout from the need to curate data, create and maintain AI infrastructures, train and maintain AI systems, develop new devices and applications, provide new expert services, etc. Most importantly, though, the promise is that our interaction with our tools and systems will be revolutionized, the same way that direct manipulation interfaces on our smart phones (Apple iPhone) revolutionized it 15 years ago. Personally, I do not fear the future. As a scientist and technologist, I am ready to embrace it.