The next new wave of Artificial Intelligence is here in the form on Chatbots — which enables end users to communicate directly with machines that are programmed to converse with humans. Several message interfaces such as Facebook Messenger are perfect avenues for deploying chatbots on already existing chat frameworks.
Chatbots provide a huge advantage over live communication channels, as they provide instant feedback to an end user without requiring human resources. Chatbots are also more powerful than search engines, as they can potentially perform various Question Answering (QA) based functions that traditional Information Retrieval (IR) based system such as search engines are unable to do. For instance the query — “how much did I spend on Amazon last month?”, requires selecting all Amazon transactions based on time period of last month, and computing the total cost. This is sophisticated reasoning that a simple search query (like google search) is not able to do.
Chatbots — like the general space of Natural Language Processing (NLP) face the issue of requiring a lot of domain dependent modeling in order to be smart and effective. For instance, Kai — which is a MasterCard chatbot can answer questions about your recent financial transactions, and aggregate that information to the user in a usable and useful manner. This chatbot needs to know about arithmetic, searching user logs, answering generic questions about MasterCard and also how to identify critical vs regular questions. This is an assimilation of quite a few areas of NLP research such as — Question Answering, Keyword Identification, Information Retrieval, Sentiment Analysis, among others.
Generic chatbots are hard to develop because of the specialized domain modeling. Each chatbot needs to be carefully developed to address questions regarding the domain that it needs to be an expert in. For instance, a weather bot needs to be aware of all keywords related the the weather, resources about how to find weather instantaneously, and the current colloquial language talking about weather. It is inconceivable to use a weather bot for another purpose (like making it into a shopping bot). Additionally, the strategy behind building a bot should be well conceived and thought out. For instance, Home Depot has a service on its website that helps people find the exact location of the object they are looking for — with the isle and bay number. Macy’s (along with IBM Watson) has tried to implement a similar service using a chatbot, but it is hard to imagine people using this location based service in a store where identifying markers are unstructured and relative. For instance — if you type for shoes, it tells you the location of shoes are compared to other products in store.
In contrast to Macy’s chatbot, Best Buy has done a great job in rolling out their digital assistant. This chatbot (which is in beta model in the BestBuy iPhone App) accepts a user query and transfers the consumer to the most likely agent that can help them with the issue.
The Best Buy’s digital assistant can also help answer select questions — such as questions about price match. Although, that is about all it can do, one can envision a smarter chatbot that can help with orders, sales etc, once the backend modeling is done for this tasks.
Chatbots will continue to remain one of the more important NLP tasks for the next few years. It is important though to have a strong NLP or AI based technology to help build out the real smarts behind the engine, and a strong fronted team that will deliver the appropriate user experience to the end user so that the chatbot is usable, smart and useful!