Contrastive Fine-tuning on Few Shot Intent Detection with Topological Intent Tree

We present a few-shot intent detection model for an enterprise's conversational dialogue system. The model uses an intent topological tree to guide the search for the user intent using large language models (LLMs). The intents are resolved based on semantic similarities between user utterances and the text descriptions of the internal nodes of the intent tree or the intent examples in the leaf nodes of the tree. Our results show that an of-the-shelf language model can work reasonably well in a large enterprise deployment without fne-tuning, and its performance can be further improved with fne-tuning as more domain-specifc data becomes available. We also show that the fne-tuned language model meets and outperforms the state-of-the-art (SOTA) results in resolving conversation intents without training classifers. With the use of a topological intent tree, our model provides more interpretability to cultivate people's trust in their decisions.