The power of chatbots is not lost on the business world. As brands focus on promoting personalized experiences, more and more intelligent chatbots are being built to engage users and improve brand image. That said it is a rarity to find a live intelligent chatbot, also called as AI chatbot. As the thought of a chatbot springs up, we know it is not a real person for sure. What we know is that chatbot brings a human touch. For that to become a reality, chatbots need to be really intelligent. The crux is not the chatbot rather it is the intelligence quotient of the chatbot that can bring the human touch.
It is the intelligence that gives power to the AI chatbot to learn from conversations and handle any and every situation that comes its way. As chatbots move into complex territories, raising the intelligence quotient becomes increasingly difficult. How to build smart chatbots and what deserves our attention?
Does the chatbot know what the user wants?
A chatbot is smart when it becomes aware of user needs. For instance, let us consider the case of a live chatbot helping a user book a room in a hotel. The user is prompted to give out the date the user has in mind to book the room. So far so good until the query ‘Are premium rooms available’ comes from the user. Now the AI chatbot must understand this specific user need to provide a relevant answer. An intelligent chatbot will understand and learn the language nuances to give a convincing answer. In the future, there will come a time where the bots will have artificial intelligence which will know what we want before we even ask it.
To cut down design complexity, it is important to ignore proactive user queries by keeping it local. But an AI chat bot is based on the human capability of self-learning and gaining information efficiently. Thus it’s imperative to make the chatbot sense natural language utterances. There are tools like IBM Watson, Api.ai, and Wit.ai to incorporate natural language capability into a chat bot.
Is the chatbot a learning champion?
If a chatbot is smart, then learning becomes a distinguishing trait of the chatbot. An intelligent chatbot is one that learns conversations all the time to improve its performance. The modules in a chatbot including user modeling modules and the natural language understanding module which can perform better by learning continuously. Machine learning(ML) algorithms and human supervisors enable the learning of the chatbot. ML techniques like reinforcement learning supervised, and unsupervised techniques can be leveraged to ensure the AI chatbot becomes a good learner.
The ability to learn is a key factor in creating an intelligent chatbot. With neural networks and deep learning, chatbots can become good learners. Learning is paramount to ensure that the chatbot recognizes patterns in data it receives and responds to user requests in the most appropriate way.
Does the chatbot know how to meet user requests?
A chatbot is primarily built to serve the user request. It is crucial for the chatbot to plan how to perform the task requested by a user. Chatbot responds to each user request by learning from the conversation so as to what the request is. Progress from one user request to another also requires planning until completion of the task. When it comes to complex tasks, chatbots must identify the action sequence to do the primary goal of the user. Planning is a sequence of actions which form conversations and include acknowledgment, questions, and information. As it learns from conversations with the users it will continue growing smarter and smarter with each conversation.
How do we determine if a chatbot is intelligent?
The AI chatbot comes with the ability to fix a goal and work autonomously to achieve that goal. This is easier said than done where identifying the goal for a specific situation is a hurdle in itself to cross. The chatbot adheres to a three-step process for realizing the goal. It is the sense-think-act cycle that can define the intelligence of a chatbot. An AI chatbot goes through this cycle to make progress towards pre-defined goals autonomously.
Ability to sense
For an AI chatbot, sensing the environment where it resides becomes a prerequisite for getting the information required to perform a task. The chatbot finds it easy to listen to what the user says than make sense what is being conveyed by the user. Take the case of a robot that you want to build. It becomes a challenge to infuse sensing power into the robot for there is a dire need to integrate the robot with most modern sensors.
Sharp to think
In simple terms, chatbot must think what to do when a user places his request. The chatbot must convert information received from a user into an understandable format and store it in a knowledge base. An AI chatbot makes a decision by leveraging pre-existing knowledge and one that it acquires continuously. Based on this decision, the chatbot takes action to achieve pre-defined goals. Use neural networks in machine learning to make the chatbot think and take actions depending on the request placed by the user.
The knowledge base influences the learning capability of the chatbot from its past conversations with users. Take the case of Siri and Google Now. Their intelligence is due to the knowledge stored internally. This knowledge base helps in learning faster, identifying relevant information and providing a response that is relevant.
Information gathered and learned guides the chatbot to decide on the relevant action. Taking decision is more about what the chatbot has to reply to a user’s request. Predictive analytics using machine learning can make the AI chatbot plan ahead about queries that would come from the user. This can make the chatbot more intelligent.
Quick to act
As the thought cycle gets over, the chatbot knows the action it has to take to respond to a user. Now, the chatbot has to act. The chatbot must now type out the reply to a specific query raised by the user. Typing out a sentence is relatively easy for a chatbot when compared to responding via its audio or video capabilities. For audio or a video chatbot, responding to the user through a suitable action becomes difficult in the way it has to sound like a human.
What do you want the chatbot to do?
Infusing the intelligent quotient into your chatbot also depends on what you want your chatbot to do. You can either make the chatbot help the user or collect information from the user. A chatbot acting as a helper is considered to be smarter than the chatbot that serves as a collector. The helper chatbot interprets what the user is saying and performs the task for the user. The intelligent chatbot could help the user buy products, seek information about cars or even book a hotel room. What are the characteristics that define the helper chatbot?
The helper chatbot is recognized by its natural language processing(NLP) and understanding power. Collector chatbots, in turn, leads the conversation with the user. They adhere to pre-defined question models and are not smart enough to respond when a user raises a query. The drive to increase the intelligent quotient of the collector chatbots depends on the intelligent platform where they are built to reside. How can we build intelligence into a collector chatbot?
A collector chatbot becomes intelligent when it responds by collecting information from the user and presenting it in the most appropriate way to serve the user’s purpose.
What is the model for the intelligent chatbot?
A chatbot based on the retrieval-based model works on the concept of predefined responses. The chatbot picks appropriate responses from the repository stacked which is based on the context and query raised by the user. Generative models built using machine translation techniques come with the ability to generate new responses right from the word go. Generative models enable longer conversations where the chatbot deals with several user queries. Though deep learning techniques are leveraged for building both these models, generative models seem to draw more power than its counterpart.
How would you want the conversation to progress?
If you do not want to limit the conversation to a single goal or intention, then open domain proves to be the right fit. In this case, the conversation can take off in different directions and topics. In turn, AI chatbot must have the knowledge to create responses for queries involving various topics. Conversations happening in social media come close to the open domain category. On social media, the conversation is not narrowed down to a single topic as the conversation goes in different directions.
When you want to limit inputs as well as outputs, closed domain comes up as the best choice. Closed domain category works well for the chatbot built to achieve specific goals. Sales support system falls into this category where the topic doesn’t veer off in other directions.
Do we foresee challenges in building intelligent chatbot?
Building an intelligent chatbot is not devoid of challenges. From making the chatbot context-aware to building the personality of the chatbot, there are challenges involved in making the chatbot intelligent.
Sensible responses are the holy grail of the chatbots. Integrating context into the chatbot is the first challenge to conquer. In integrating sensible responses, both the physical context as well as linguistic context must be integrated. For incorporating linguistic context, conversations are embedded into a vector, which becomes a challenging objective to achieve. While integrating contextual data, location, time, date or details about users and other such data must be integrated with the chatbot.
Achieving coherence is another hurdle to cross. The chatbot must be powered to answer consistently to inputs that are semantically similar. For instance, an intelligent chatbot must provide the same answer to queries like ‘Where are you from’ and ‘where do you reside’. Though it looks straightforward, incorporating coherence into the model is more of a challenge. The secret is to train the chatbot to produce semantically consistent answers.
How is the chatbot performing?
The answer to this query lies in measuring whether the chatbot performs the task that it has been built for. But, measuring this becomes a challenge as there is reliance on human judgment. Where the chatbot is built on an open domain model, it becomes increasingly difficult to judge whether the chatbot is performing its task. There is no specific goal attached to the chatbot to do that. Moreover, researchers have found that some of the metrics used in this case cannot be compared to human judgment.
In some cases, reading intention becomes a challenge. Take generative systems for instance. They provide generic responses for several of user inputs. The ability to produce relevant responses depends on how the chatbot is trained. Without being trained to meet specific intentions, generative systems fail to provide the diversity required to handle specific inputs.
Plan to use NLP and machine learning?
Another factor that deserves attention is the plan to leverage NLP or machine learning for building the intelligent chatbot. In the case of natural language processing, it is about finding answers by parsing language into intent, entities, agents, actions, and contexts. With NLP reckoned as the driving force, NLP platforms like WIT, API, and LUIS can be leveraged to build an intelligent chatbot.
While you plan to leverage machine learning to create your own NLP, you must decide upon the model prior to building the intelligent chatbot. It is important to weigh generative and retrieval-based model, open and closed domains to create the intelligent chatbot that you have in mind.
Is an intelligent platform ‘the’ alternative?
Building a smart chatbot is one school of thought. Building a chatbot on an intelligent platform is altogether a different one. Today, several of successful chatbots including x.ai and Google assistant have been built on intelligent platforms. In this scenario, the platform becomes the intelligent agent, and the chatbot becomes a sensor for this intelligent agent.
The intelligent platform works to find out the goal, collect user information, process, store and convert information to realize the goal. Then the challenge is not about infusing intelligence into a chatbot but creating an intelligent platform. The focus must fall on ways to define the goal and factor sense-think-act capability into the platform.
For now, the chatbot imperative is to meet user-centric tasks. For that to happen, the chatbot must be smart and knowledgeable. The chat to build a smart chatbot gets chattier when significant elements surrounding the building process make an entry. As we look into the future, intelligent chatbots will be built to rule the world of connections.