Blog

Developer
Concepts
Use Cases
Product
Spotlight

Conversation Understanding: Open Domain vs. Closed Domain

In short, a closed domain system is built to only understand specific kinds of conversations, like a sales chatbot; whereas an open domain system is built to understand any conversation in any domain. Choosing between them depends on the scope and complexity of the conversation it’s expected to follow, which can be human to machine (H2M) or human to human (H2H).

What’s the difference between open domain and closed domain conversation understanding?

A closed domain system, also known as domain-specific, focuses on a particular set of topics and has limited responses based on the business problem. For example, a pizza delivery chatbot can only help you place, track, or cancel an order.

This straight-shooting conversation is like bumping into an acquaintance: you know they’re going to ask about your job, your family, and maybe comment on the weather. You have a predefined response to each question and the objective is simply to satisfy their inquiries and move on.

On the other hand, an open domain system is expected to understand any topic and return relevant responses. For example, Meena — Google’s open domain bot that can “chat about anything.”

Open domain is more like talking to a friend where the conversation can go in any direction — from what series you’re binging to why your client demo stopped working exactly two seconds after sending it. This free-flowing conversation has no defined objective, so your responses need to adapt to whatever information you’re given.

How do open domain and closed domain systems work?

Open domain and closed domain are two very different approaches to building a conversation understanding system (CUS). A CUS models the context of a conversation so that the same model can be applied to different conversations in different use cases.

This CUS can be applied to either human to machine (H2M) or human to human (H2H) conversations. H2M involves a person typing or speaking to a conversational agent, like a chatbot or voice assistant. On the other hand, H2H can be between two or more people having a conversation in any context, like a business meeting or gossiping with a friend.

To handle these conversations, AI models typically have to detect intent, identify entities for context (e.g. names, locations, dates, etc.), and then generate a response. When talking about H2M conversations, like with a sales bot, there are two main ways for the AI to generate a response:

Retrieval-based system: When it’s the machine’s turn to respond, the model reaches into its repository of predefined responses and picks the most relevant answer. This system can be simply rule-based or rely on machine learning to retrieve the best response. Retrieval systems can only use the text they are given and can’t generate new answers.

Generative-based system: In this case, responses are actually generated from scratch depending on the given conversation history (i.e. context). This system can handle both common and unforeseen questions, making it appear more human-like and better for longer conversations. However, this savvy response system also increases implementation complexity.

For H2H conversations, either a closed domain or open domain system can work closely with natural language understanding (NLU) to infer context and respond appropriately. However, it’s tough to get right, since these conversations are highly unstructured and can be contextually ambiguous.

For example, think of your typical brainstorming meeting. It has no fixed questions or limits on the vocabulary and the conversation can run in any direction. There can be arguments, slang, hypotheticals, and references to previous topics. Such an unpredictable conversation would be tough for a closed, domain-specific system to handle.

On the other hand, whether the conversation happens in real time or has been recorded, a fine-tuned open domain CUS can listen in, understand what’s being said, and then pick out the most important bits for later — just like a human would.

Which conversation understanding system should you use?

Choosing the best CUS largely depends on the nature of the conversation you expect it to understand. For H2M or H2H conversations, deciding between an open or closed domain system boils down to the scope of the conversations you’re working with.

When the scope is limited, like a customer care call with a fixed set of questions, a closed domain CUS can get the job done. But when the scope is unlimited, like with H2H meetings, e-learning lectures, or customer support; an open domain CUS is better equipped to capture the right context and use it to perform conversation understanding tasks (CUT), such as:

  • Noting conclusions, action items, and follow-ups
  • Pulling questions raised during the conversation
  • Listing important information, like topics and open issues
  • Automatically suggesting the next step, like set a meeting.

At a broader level, there are a few more factors to consider:

  • Time-to-market: Closed domain is ideal when you have a specific problem to solve (e.g. a customer service chatbot) and need to provide business value pronto. But if you’re facing several use cases, like transcribing H2H sales calls and pulling action items from team meetings, it’s quicker to calibrate an open domain CUS to specific tasks rather than build a new closed domain CUS for each one.
  • Generalization: The narrower the scope of the conversations, the easier it will be to build a CUS for it. As you might guess, short and straight-forward H2M conversations are prime closed domain territory. If you need to generalize the conversation understanding across different domains, an open CUS will be easier than stuffing individual closed CUS with tremendous amounts of domain-specific data.
  • Scalability: An open domain CUS is much easier to scale since you can use the same system across different use cases. With a closed domain CUS, you may find yourself juggling multiple AI systems for different domains and tasks, which could become tedious and expensive in the long run.

When it comes to building the actual CUS, you can either outsource it or do it yourself with an in-house team and/or use third-party APIs to get it to market faster. Whichever you choose, it’s worth getting familiar with different CU systems so you can leverage the right one to squeeze the most value out of every conversation.

Additional reading

There’s a lot more to learn about the exciting realm of open and closed domain conversation understanding. Here are some handy resources you can dig into:

Sign up for Symbl updates