Dialogue flow using semantic simplexes

A computer-implemented method for providing a computer-implemented medical diagnosis via a computer, the computer being configured to generate questions to ask to a user when performing the diagnosis, the computer comprising a memory, and the method comprising:

• constructing an energy landscape as a semantic space for the questions,
• wherein the semantic space is represented by at least one semantic simplex defining an embedded space;
• embedding nodes of a probabilistic graphical model (PGM) in said embedded space,
• wherein the PGM comprises probability distributions and relationships between evidence and diseases, and
• wherein the evidence comprises symptoms and risk factors, the PGM further comprising evidence nodes relating to risk factors and symptoms and disease nodes;
• receiving, at the computer from a user device, an input from the user comprising at least one symptom of the user;
• providing the at least one symptom as the input to a medical model comprising said PGM, the medical model being stored in the memory,
• wherein the at least one symptom is provided to the PGM by activating an evidence node in the PGM corresponding to the at least one symptom,
• performing Bayesian inference utilizing an inference engine on said PGM to calculate an expected information gain of activating each node of a set of further evidence nodes;
• retrieving, from the memory, embedded evidence nodes of the PGM,
• wherein the embedded evidence nodes correspond to evidence nodes of the PGM that have been embedded in said semantic space;
• calculating a cost, for each of the embedded evidence nodes of the set of further evidence nodes, of moving through the energy landscape of the semantic space from an embedded state vector in the semantic space to an embedded evidence node,
• wherein the embedded state vector is determined from a previous question asked by the computer to the user;
• ranking said set of further evidence nodes using, for each node, said calculated expected information gain penalized by the calculated cost of moving through the energy landscape of the semantic space from the embedded state vector;
• using said ranked further evidence nodes to select a next question to be provided by the computer to the user;
• and
• outputting said next question to the user through the user device.