US 12,450,464 B2
Guided dialogue using language generation neural networks and search
Geoffrey Irving, London (GB); Amelia Marita Claudia Glaese, London (GB); Nathaniel John McAleese-Park, London (GB); and Lisa Anne Marie Hendricks, London (GB)
Assigned to GDM Holding LLC, Mountain View, CA (US)
Filed by GDM Holding LLC, Mountain View, CA (US)
Filed on Sep. 20, 2023, as Appl. No. 18/471,257.
Claims priority of provisional application 63/408,430, filed on Sep. 20, 2022.
Prior Publication US 2024/0104336 A1, Mar. 28, 2024
Int. Cl. G06N 3/006 (2023.01); G06F 40/284 (2020.01); G06F 40/35 (2020.01); G06N 3/0455 (2023.01); G06N 3/092 (2023.01)
CPC G06N 3/006 (2013.01) [G06F 40/284 (2020.01); G06F 40/35 (2020.01); G06N 3/0455 (2023.01); G06N 3/092 (2023.01)] 30 Claims
OG exemplary drawing
 
1. A method, implemented by one or more computers, the method comprising:
initializing a context input that represents a context for a dialogue between a user and the one or more computers, and,
at one or more of a plurality of dialogue update iterations:
receiving a natural language request from the user;
updating the context input to include the natural language request;
processing the context input using a first trained language generation neural network to generate a set of one or more first natural language responses without using a search system interface, wherein the first trained language generation neural network is a first auto-regressive neural network that includes a first succession of first self-attention neural network layers, and wherein each first self-attention neural network layer applies an attention mechanism over a first attention layer input for the first self-attention neural network layer to generate an attention layer output for each element of the first attention layer input;
generating a set of one or more search queries from the natural language request;
providing the one or more search queries to a search system using the search system interface, wherein the search system is external to the first trained language generation neural network;
receiving, for each search query, a set of one or more search results from the search system via the search system interface;
determining, from the context input and for each of the search results, a supported context input that includes content from one of the search results;
for each supported context input that includes content from one of the search results, processing the supported context input using the first trained language generation neural network to generate a set of one or more second natural language responses, wherein each second natural language response includes a response to the natural language request including supporting evidence that comprises content from the one of the search results obtained by providing the one or more search queries to the search system using the search system interface;
processing (i) at least a portion of the context input that includes the natural language request received by the user, (ii) the one or more first natural language responses generated without using the search system interface by the first trained language generation neural network in response to the context input, and (iii) the one or more second natural language responses generated by the first trained language generation neural network in response that each include a response to the natural language request including supporting evidence that comprises content from the one of the search results obtained by providing the one or more search queries to the search system using the search system interface in the content from the one using a trained response selection neural network to generate an output that defines a selection between the one or more first natural language responses and the one or more second natural language responses, wherein the trained response selection neural network is a second auto-regressive neural network that includes a second succession of second self-attention neural network layers, and wherein each second self-attention neural network layer applies an attention mechanism over a second attention layer input for the second self-attention neural network layer to generate an attention layer output for each element of the second attention layer input;
selecting, based on the output of the trained response selection neural network and as a natural language reply, one of the one or more first natural language responses generated by the first trained language generation neural network in response to the context input and the one or more second natural language responses generated by the first trained language generation neural network in response to each supported context input;
providing the natural language reply to the user as a response to the natural language user request; and
updating the context input to include a representation of the natural language reply for the next dialogue update iteration.