Final: Automatic design of conversational models from observation of human-to-human conversation

Currently used conversation models (or dialog models) are mostly hand designed by data analysts as a conversation graph consisting of the system’s prompts and the user’s answers. The advanced conversation models are based on large language models fine-tuned on the dialog task, and still require significant amounts of training data. These models produce surprisingly fluent outputs but are not trustable because of hallucination (which can produce unexpected and wrong answers), and their adoption in commerce is limited.

Our goal is to explore ways to design conversation models in the form of finite state graphs semi-automatically or fully automatically from an unlabeled set of audio or textual training dialogs. Words, phrases, or user turns can be converted to embeddings using (large) language models trained specifically on conversational data. These embeddings represent points in a vector space and carry semantic information. The conversations are trajectories in the vector space. By merging, pruning, and modeling the trajectories, we can get dialog model skeleton models. These models could be used for fast data content exploration, content visualization, topic detection, and topic-based clustering, speech analysis, and mainly for much faster and cheaper design of fully trustable conversation models for commercial dialog agents. The models can also target some specific dialog strategies – the fastest way to reach a conversation goal (to provide useful information or sell a good or entertain users for the longest time). One promising approach to building a conversational model from data is presented in. Variational Recurrent Neural Networks are trained to get discrete embeddings with a categorical distribution. The categories are conversation states. Then a transition probability matrix among states is calculated, and low probabilities are pruned out to get a graph.

Mots clés : deep nets ia information retrieval informatique jsalt linear algebra nlp workshop