Capital One has filed a patent for a computer-implemented method that uses machine classifiers to generate responses in multi-turn dialogues. The method involves training a generator and a discriminator using adversarial bootstrapping, with the generator utilizing a hierarchical recurrent encoder-decoder network and the discriminator employing a bi-directional recurrent neural network. The responses are ranked based on discriminator feedback, and the optimal response is selected and transmitted. GlobalData’s report on Capital One Financial gives a 360-degree view of the company including its patenting strategy. Buy the report here.
According to GlobalData’s company profile on Capital One Financial, virtual banking assistant was a key innovation area identified from patents. Capital One Financial's grant share as of June 2023 was 1%. Grant share is based on the ratio of number of grants to total number of patents.
The patent filed is for a computer-implemented method for generating responses
A recently filed patent (Publication Number: US20230206009A1) describes a computer-implemented method for generating responses in a conversation. The method involves receiving communication data that includes multiple dialog sequences. Using a generative adversarial network, a generator generates a variety of responses based on context embedding and word embedding of the communication data. Each response includes at least one keyword selected through maximum likelihood estimation. The responses are then ranked by at least one discriminator, also part of the generative adversarial network, based on the context embedding and word embedding. An optimal response is selected by the discriminator from the ranked responses, and this optimal response is transmitted.
The patent also mentions that the generator and the discriminator share the context embedding and word embedding. Additionally, an encoder is used to generate the word embedding of the communication data. The generator further generates initial autoregression data and initial teacher forcing data based on conversation data from training data. The discriminator then determines its accuracy based on the initial autoregression data and initial teacher forcing data. If the accuracy falls below a threshold, the discriminator is trained, and the generator is retrained using a teacher forcing loss function. If the accuracy is above the threshold, the generator is retrained using both the teacher forcing loss function and an autoregressive loss function. The trained generative adversarial network is stored for future use.
The discriminator also assigns a discriminator score to each response, indicating the statistical similarity between the response and an anticipated ground truth response to the communication data. The responses are further ranked based on these discriminator scores. Alternatively, the discriminator can assign a discriminator score indicating the probability that a response is a ground truth response to the current prompt indicated in the communication data.
The patent also describes a computing device for generating responses, which includes a processor and memory storing instructions for executing the method described above. Furthermore, a non-transitory machine-readable medium is mentioned, which stores instructions for performing the steps of the method when executed by one or more processors.
In summary, the patent outlines a computer-implemented method and system for generating responses in a conversation using a generative adversarial network. The method involves receiving communication data, generating responses based on context and word embeddings, ranking the responses, selecting an optimal response, and transmitting it. The generator and discriminator share embeddings, and the system includes an encoder, training processes, and scoring mechanisms for ranking the responses.