US 12,282,743 B2
Autonomous conversational AI system without any configuration by a human
Amol Kelkar, Redmond, CA (US); Nikhil Varghese, San Francisco, CA (US); Chandra Khatri, San Francisco, CA (US); Utkarsh Mittal, Foster City, CA (US); Nachiketa Rajpurohit, San Mateo, CA (US); Peter Relan, Atherton, CA (US); and Hung Tran, Foster City, CA (US)
Assigned to GICRM AI LLC, Palo Alto, CA (US)
Filed by GICRM AI LLC, Palo Alto, CA (US)
Filed on Jan. 6, 2022, as Appl. No. 17/570,281.
Claims priority of provisional application 63/135,840, filed on Jan. 11, 2021.
Prior Publication US 2023/0274095 A1, Aug. 31, 2023
Int. Cl. G06F 17/00 (2019.01); G06F 40/247 (2020.01); G06F 40/289 (2020.01); G06F 40/49 (2020.01); H04M 3/527 (2006.01)
CPC G06F 40/49 (2020.01) [G06F 40/247 (2020.01); G06F 40/289 (2020.01); H04M 3/527 (2013.01)] 22 Claims
OG exemplary drawing
 
1. A method for automatically generating a configuration for an autonomous conversational artificial intelligence (ACAI) system, the method comprising:
receiving a conversation log comprising interactions between at least two actors;
processing the conversation log to normalize the conversation log into a canonical representation;
generating a first configuration for a Natural Language Understanding model of the ACAI system, by:
utilizing a first generative language model fine-tuned on a domain-specific dataset to annotate user utterances in the conversation log with turn-level auto-intents and auto-responses;
annotating each conversation with conversation-level auto-topics and auto-subtopics using a second generative language model, wherein the second generative language model is fine-tuned on a dataset containing similar topic and subtopic annotations from the same domain; and
selecting representative conversations based on the frequency of auto-topic and auto-subtopic pairs to maximize coverage of interactions between actors;
generating a second configuration for a Dialog Management model of the autonomous conversational system, by:
converting the annotated conversations into a graph of sentence-level auto-intents and turn-level auto-responses using a third generative language model fine-tuned on a dataset with sentence-level intent annotations;
identifying and aligning matching parts across conversation flows using a multi-sequence subsequence alignment algorithm; and
selecting flows with user intents and system responses above a specified frequency threshold to simplify the second configuration; and
automatically configuring the ACAI system using the generated first and second configurations, wherein the ACAI system, once configured, is trained to mimic behaviors based on learned conversation patterns.