Presented by: Sanaz Bahargam – Applied Scientist at Amazon Lab126

Goal-oriented dialog systems enable users to complete specific goals like booking a flight, ordering food, or receiving weather information. In most cases, the dialog systems consist of multiple ML models trained on labeled data to perform natural language understanding, state tracking, policy learning, natural language generation, and slot filling. A grand challenge in this scenario is getting labeled data for each domain, as the annotated data for dialogs is scarce and annotation is time-consuming and expensive. In this talk, I address how we can overcome this challenge by generating annotated data via a novel dialogue simulator based on a few seed dialogues and specifications of APIs and entities provided by the developer.

In addition, I will discuss different components of the dialog system such as dialog context encoder, NER, action prediction, and argument filling and how we train each model. Finally, I will show the technique can significantly reduce developer burden while making robust experiences.

свернуть