Balancing Control and Flexibility in a Task-oriented Chatbot: an Industry Use Case

When building a task-oriented chatbot there is a fine line between having enough control over the system’s output and being flexible in generating answers about questions that go outside of the system’s knowledge base.

In this work we present an industry use case where a task-oriented chatbot helps a human operator to assemble a compressor. The agent needs to deal with similar questions that can be asked during each step of the process or about each assembly part. Examples of such questions are “Where can I find this part?”, “Where should I place this part?”, or “Do I need a tool for this?”. Rather than having to add all types of different questions to the training data, we opted for a solution with general training data making use of slots for the different steps and assembly parts in the process in combination with a knowledge base containing all the relevant information for the assembly process. Linking the slot values to the names of the different levels in the knowledge base allows us to track the state of the assembly process. It also enables the chatbot to answer adequately to a question like “What’s next?”, which on its own does not contain any information about the current state of the assembly process. This part of the system has high control over the output by combining intent prediction (text classification) with template-based answers where specific information from the knowledge base is inserted in the templates’ slots.

However, besides having high control over the output, the system should be flexible enough to deal with questions that go outside of the scope of the knowledge base. For example, inexperienced operators might ask about specific assembly parts or tools. Such information is typically not present in the description of the assembly process and seems superfluous to add to the knowledge base. Instead, we integrated ChatGPT functionality to generate such answers. This results in needing less training data and being flexible enough to recognize new assembly parts or tools that would be needed for a new assembly process.

A final objective of this study is to test the transferability of the system to deal with other assembly processes by just using new knowledge bases.