Conversation between two humans is often more natural and effortless compared to the conversation between humans and virtual assistants in their current state. The primary reason behind the same is the cognitive ability of humans to remember historical interactions. This allows humans to start from the common ground of understanding and have more context-aware conversations. However, this cognitive ability of remembering and understanding historical context is very hard to establish in virtual assistants even though it is essential for a good conversational experience. Furthermore, the term “contextual conversations” is used ambiguously while evaluating the capability of a virtual assistant. In this blog, I will elaborate on the meaning of contextual conversations with respect to virtual assistants and break the term context into concrete measurable components.
How to define context in Virtual Assistants?
At Haptik, we primarily break down context into four components and each component is explained below along with examples.
- Conversation specific context
- User context
- Domain-specific context
- Business context
Conversation specific context
- Definition - Conversation-specific context consists of all the information exchanged in previous dialogues between the user and virtual assistant. A few important operations that virtual assistants should support in order to retain conversation context are mentioned below
- Memory Storage - Basic information such as Intent of the user, Entities collected during the conversation, and other relevant metadata from historical messages needs to be stored and should be accessible at any point in the conversation.** This feature is limited by the maximum number of historical messages and time period of history supported by the conversational AI platform.
- Anaphora Resolution - Messages such as “Book it for me”, “Where is it?” requires resolution of the third-person pronoun “it” in order to send the correct response
- Ellipsis Resolution - Messages such as “why?”, “Are you sure?”, “Go ahead” requires attending to relevant previous messages.
- General Sequence Understanding - Messages such as “Change it to tomorrow instead”, “Sorry I meant debit card”, “98334305XX” need a virtual assistant to have the capability to overwrite relevant information from the past.
- Application - Conversation-specific context enables experience which would allow customers to switch across intents and progress towards their goal without having to repeat the same information again during their journey.
- Definition - User context consists of user information such as name, contact details, last order placed, seat preference in flight, etc.
- Application - Offering personalized experience and proactive assistance are primary applications of User context.
- Definition - Domain-specific context consists of knowledge and common facts which the customer assumes that the virtual assistant will understand during a conversation. It is the hardest to achieve and is a major differentiator why many customers prefer talking to humans over virtual agents today.
- Application - Domain knowledge of virtual assistant support plays an important role in ensuring that customers need not explicitly specify every detail in the request.
- Definition Business context comprises of business logic or data that virtual agent needs to access in order to provide the correct response.
- Application - Virtual assistants can understand rules for loan eligibility or predict the next purchase by accessing third-party data or logic at different stages.
Achieving high accuracy on Context components still remains an overarching goal of the AI community and we at Haptik are committed to pushing the boundaries of the same. The capability to support the above functionalities sits at the heart of building personalized and natural conversations and I hope this article helps you establish a clear definition of Contextual conversations. Get started with the Haptik Platform to check out context retention capability and build Intelligent Virtual Assistants.