Intelligent Dialogue Technologies

Authored by Debashis Guha, Associate Professor and Director – Machine Learning

ABuilding computer systems that can carry on a dialogue with humans has been one of the central topics in the field of AI from its very inception. Alan Turing’s 1950 proposal for a question and answer imitation game to test for digital intelligence, known as the TuringTest, is a foundational concept of the field. The quest for practical implementations of the Turing Test led to early dialogue systems such as ELIZA, PARRY, and RACTER.

In recent years, rapid advances in Natural Language Processing, especially techniques using Deep Learning, have led to the development of practical tools for building dialogue systems that have many practical applications. Principal among these practical applications are enterprise chatbots used by businesses as first-line responders or communicators for marketing and customer support. Other major fields of application include virtual assistants that help in performing daily tasks, natural language information retrieval, and digitalcompanionship for the elderly.

All conversational AI systems have three principal functions, question answer dialogue, task assistance dialogue, and general-purpose dialogue.

Question and Answer Dialogue

Question Answer (QA) agents provide a natural language interface to databases and knowledge bases that enable users to search using natural language queries. These agents are more user friendly than SQL-like querying systems, and they can be used interactively to retrieve information. Modern QA agents are built using deep learning and they use vector embedding for semantic inference.

Search providers such as Google, Microsoft Bing, and Baidu already offer QA interfaces, albeit somewhat rudimentary ones, to their search engines. For instance, Bing QA is an extension of the Microsoft Bing search engine and offers a text-based QA agent that generates natural language responses to natural language queries entered into the Bing search box.

Task Assistance Dialogue

The second major function of conversational AI is to assist users in performing a well specified task, such as making a flight reservation, or filling out a form etc. The task orientation of these systems differentiates them from QA agents that are used to retrieve information. Such task-assistance agents have a well specified goal and are usually built for a specific domain. Task assistance agents use deep learning and deep reinforcement learning.

Personal assistants such as Amazon’s Alexa, Apple’s Siri, Google’s Assistant and Microsoft’s Cortana are available both on smartphones and laptops to help users navigate and complete daily tasks. These assistants can also be used to retrieve information from a knowledge base residing on a central server.

General Purpose Dialogue

Another very important function of a dialogue agent is the ability to carry on a general-purpose conversation. Such a conversation has no fixed goal, although sometimes its overall purpose may be to shrink the domain of the conversation, in order to carry out an information retrieval step or a task completion step. General purpose dialogue is usually based on a model trained on past data, using deep recurrent networks for sequence to sequence prediction.

All personal assistants such as Alexa, Cortana, Google, and Siri also function as general-purpose dialogue systems.

Development Platforms

The most widely used platforms that are used to develop virtual assistants, chatbots and other dialogue agents are Amazon’s Alexa SkillsKit, Facebook’s MessengerPlatform, IBM’s WatsonAssistant, Google’s Dialogflow, and Microsoft’s AzureBot Service.

Alexa Skills Set (ASK) a collection of self-service APIs, tools, documentation, and code samples. ASK powers Amazon’s devices such Echo and FireTV and some third-party devices incorporating Alexa.

Facebook’s Messenger Platform is meant for handling customer inquiries. When a customer sends a message to a business on Messenger, the Facebook server communicates with a messaging app on the business server, and using an API, the messaging app can respond to the message.

IBM’s Watson Assistant is an AI that can be used to build, train, and deploy conversational interactions into any app, device, or communication channel. Watson Assistant can use the current context to decide whether to retrieve information from a knowledge base, or whether to ask for clarifications, and when to hand off to a human expert. It can be deployed on the cloud or locally.

Google’s Dialogflow can be used to build voice and text-based conversational interfaces, such as voice assistants and chatbots. It is optimises for Google Assistant, and can also connect with Amazon Alexa and Facebook Messenger. Dialogflow incorporates deep learning and runs on Google Cloud.

Microsoft’s Azure Bot Service can be used to build a Q&A bot or a virtual assistant, or a chatbot. It has an open source SDK and tools that connect to many popular channels and devices, and it is integrated with Azure Cognitive Services, Microsoft’s family of AI services and cognitive APIs

Conclusions

Conversational AI that can carry on dialogues for question answering, task completion assistance, and general purpose conversation are becoming widely available and offered as integrated development platforms by leading vendors such as Amazon, Facebook, IBM, Google, and Microsoft. Most of these are based on deep learning, combined with some hand-crafted tools.

Some of the future trends in the development and research in conversational AI include the incorporation of common knowledge and emotional context into conversations, the capacity for providing explanations and clarifications for answers and task assistance, and the ability to adapt smoothly to the changing flow of a dialogue.