Ever wondered if the agent you are chatting with online is a human or a robot? The answer would be the latter for an increasing number of industries. Conversational agents or chatbots are being employed by organizations as their first-line of support to reduce their response times.
The first generation of bots were not too smart, they could understand only a limited set of queries based on keywords. However, commoditization of NLP and machine learning by Wit.ai, API.ai, Luis.ai, Amazon Alexa, IBM Watson, and others, has resulted in intelligent bots.
What are the different chatbot platforms?
There are many platforms out there which are easy to use, like DialogFlow, Bot Framework, IBM Watson etc. But most of them are closed systems, not open source. These cannot be hosted on our servers and are mostly on-premise. These are mostly generalized and not very specific for a reason.
DialogFlow vs. RASA
- Formerly known as API.ai before being acquired by Google.
- It is a mostly complete tool for the creation of a chatbot. Mostly complete here means that it does almost everything you need for most chatbots.
- Specifically, it can handle classification of intents and entities. It uses what it known as context to handle dialogue. It allows web hooks for fulfillment.
- One thing it does not have, that is often desirable for chatbots, is some form of end-user management.
- It has a robust API, which allows us to define entities/intents/etc. either via the API or with their web based interface.
- Data is hosted in the cloud and any interaction with API.ai require cloud related communications.
- It cannot be operated on premise.
Rasa NLU + Core
- To compete with the best Frameworks like Google DialogFlow and Microsoft Luis, RASA came up with two built features NLU and CORE.
- RASA NLU handles the intent and entity. Whereas, the RASA CORE takes care of the dialogue flow and guesses the “probable” next state of the conversation.
- Unlike DialogFlow, RASA does not provide a complete user interface, the users are free to customize and develop Python scripts on top of it.
- In contrast to DialogFlow, RASA does not provide hosting facilities. The user can host in their own sever, which also gives the user the ownership of the data.
What makes RASA different?
Rasa is an open source machine learning tool for developers and product teams to expand the abilities of bots beyond answering simple questions. It also gives control to the NLU, through which we can customize accordingly to a specific use case.
Rasa takes inspiration from different sources for building a conversational AI. It uses machine learning libraries and deep learning frameworks like TensorFlow, Keras.
Also, Rasa Stack is a platform that has seen some fast growth within 2 years.
- Intent: Consider it as the intention or purpose of the user input. If a user says, “Which day is today?”, the intent would be finding the day of the week.
- Entity: It is useful information from the user input that can be extracted like place or time. From the previous example, by intent, we understand the aim is to find the day of the week, but of which date? If we extract “Today” as an entity, we can perform the action on today.
- Actions: As the name suggests, it’s an operation which can be performed by the bot. It could be replying something (Text, Image, Video, Suggestion, etc.) in return, querying a database or any other possibility by code.
- Stories: These are sample interactions between the user and bot, defined in terms of intents captured and actions performed. So, the developer can mention what to do if you get a user input of some intent with/without some entities. Like saying if user intent is to find the day of the week and entity is today, find the day of the week of today and reply.
Rasa has two major components:
- RASA NLU: a library for natural language understanding that provides the function of intent classification and entity extraction. This helps the chatbot to understand what the user is saying. Refer to the below diagram of how NLU processes user input.
- RASA CORE: it uses machine learning techniques to generalize the dialogue flow of the system. It also predicts next best action based on the input from NLU, the conversation history, and the training data.
This diagram shows the basic steps of how an assistant built with Rasa responds to a message:
The steps are as follows:
- The message is received and passed to an Interpreter, which converts it into a dictionary including the original text, the intent, and any entities that were found. This part is handled by NLU.
- The Tracker is the object which keeps track of conversation state. It receives the info that a new message has come in.
- The policy receives the current state of the tracker.
- The policy chooses which action to take next.
- The chosen action is logged by the tracker.
- A response is sent to the user.
Areas of application
RASA is all one-stop solution in various industries like:
- Customer Service: broadly used for technical support, accounts and billings, conversational search, travel concierge.
- Financial Service: used in many banks for account management, bills, financial advices and fraud protection.
- Healthcare: mainly used for fitness and wellbeing, health insurances and others
As any machine learning developer will tell you, improving an AI assistant is an ongoing task, but the RASA team has set their sights on one big roadmap item: updating to use the Response Selector NLU component, introduced with Rasa 1.3. “The response selector is a completely different model that uses the actual text of an incoming user message to directly predict a response for it.”
About the Author –
Deepti is an ML Engineer at Location Zero in GAVS. She is a voracious reader and has a keen interest in learning newer technologies. In her leisure time, she likes to sing and draw illustrations.
She believes that nothing influences her more than a shared experience.