Introduction to Large Language Models (LLMs)

Large Language Models (LLMs) are a hot topic in the world of AI research, particularly in Dialog State Tracking (DST) where they have shown great potential. One area where LLMs excel is in-context learning, where they retrieve and add similar examples to a prompt to improve performance. However, this method requires access to labeled training data, which can be difficult and expensive to obtain.

The Need for Synthetic Data Generation

Zero-shot learning is a method that doesn’t require any training data, but it falls short compared to few-shot learning, which does use some training data. This leads to the question of whether synthetic data could be generated efficiently to enable few-shot prompting for any dialogue schema.

Introducing SynthDST

SynthDST is a data generation framework designed specifically for DST that utilizes LLMs. This approach only needs the dialogue schema and a few carefully crafted dialogue templates to create natural, coherent dialogues with DST annotations. When used for few-shot learning, SynthDST has been shown to improve Joint Goal Accuracy by 4-5% over the zero-shot baseline on MultiWOZ 2.1 and 2.4. In fact, our few-shot learning method achieves almost 98% of the performance of using human-annotated training data.