Advances in technology for implementing artificial neural networks for natural language processing have greatly increased their performance. At the same time, these advances have greatly increased the computing resources and time needed to train these networks. Our patented method for implementing neural networks reduces training time and compute costs by allowing for greater parallelization of computations. The trained neural network may be applied to a wide variety of applications, such as performing speech recognition, determining a sentiment of text, determining a subject matter of text, answering a question in text, or translating text to another language.
Discovering and prioritizing conversation patterns for automation can be a daunting challenge. Conversations may have varying flows (e.g., different ordering of messages in exchanging relevant information) and language, so automated techniques may not perform well in extracting valuable information. Our patented method for discovering and distilling conversational patterns overcomes this challenge. We condense a large collection of possible conversational flows to a much smaller number of representative flows, compute the distance between them, and cluster them using the distances. Representative conversations can then be presented to a user to review and prioritize automation.