# US 12,008,329 B2 · Systems and methods for generating dynamic conversational responses through aggregated outputs of machine learning models > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,008,329 B2 - Document ID: US12008329B2 - Status: Granted - Kind: B2 - Application number: 18/346,177 - Filed: 2023-06-30 - Published: 2023-11-02 - Granted: 2024-06-11 - Earliest claimed: 2020-09-23 - Assignees: Capital One Services, LLC - Applicants: Capital One Services, LLC - Inventors: Minh Le - Web page: https://patentgrep.com/patent/US12008329B2 ## CPC - G06F 40/35: Discourse or dialogue representation - G06N 3/044: Recurrent networks, e.g. Hopfield networks - G06N 3/045: Combinations of networks - G06N 3/0455: Auto-encoder networks; Encoder-decoder networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/088: Non-supervised learning, e.g. competitive learning - G06N 3/09: Supervised learning - G06N 3/096: Transfer learning - G06N 3/048: Activation functions - G06N 3/084: Backpropagation, e.g. using gradient descent - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract Methods and systems are described herein for generating dynamic conversational responses. For example, dynamic conversational responses may facilitate an interactive exchange with users. Therefore, the methods and systems used specialized methods to enriched data that may be indicative of a user's intent prior to processing that data through the machine learning model, as well as a specialized architecture for the machine learning models that take advantage of the user interface format. ## Claims ### Claim 1 A system for generating dynamic conversational responses through aggregated outputs of machine learning models, the system comprising: storage circuitry configured to store: a first machine learning model, wherein the first machine learning model is trained using a multi-class cross entropy loss function; and a second machine learning model, wherein the second machine learning model is trained using a binary cross entropy loss function; control circuitry configured to: receive a first user action during a conversational interaction with a user interface; generate a first output from the first machine learning model; generate a second output from the second machine learning model; determine a third output based on a weighted average of the first output and the second output; and select a subset of the dynamic conversational responses from a plurality of dynamic conversational responses based on the third output; and input/output circuitry configured to: generate, at the user interface, the subset of the dynamic conversational responses during the conversational interaction. ### Claim 2 A method for generating dynamic conversational responses through aggregated outputs of machine learning models, the method comprising: receiving a first user action during a conversational interaction with a user interface; generating a first output from a first machine learning model based on the first user action, wherein the first machine learning model is trained using a multi-class cross entropy loss function; generating a second output from a second machine learning model based on the first user action, wherein the second machine learning model is trained using a binary cross entropy loss function; determining a third output based on a weighted average of the first output and the second output; selecting a subset of dynamic conversational responses from a plurality of dynamic conversational responses based on the third output; and generating, at the user interface, the subset of dynamic conversational responses during the conversational interaction. ### Claim 3 (depends on 2) The method of claim 2, wherein the first output comprises a first plurality of probabilities that summed to one, wherein each of the first plurality of probabilities corresponds to a respective user intent. ### Claim 4 (depends on 2) The method of claim 2, wherein the second output comprises a second plurality of probabilities that summed do not sum to one, wherein each of the second plurality of probabilities corresponds to a respective user intent. ### Claim 5 (depends on 2) The method of claim 2, further comprising determining a first feature input for the first machine learning model, wherein the first feature input comprises a matrix, and wherein the first output corresponds to a prediction based on a column of the matrix and the second out-put corresponds to a row of the matrix. ### Claim 6 (depends on 2) The method of claim 2, wherein the first machine learning model comprises training a single classifier per class, wherein samples of the class are positive samples and all other samples are negative samples. ### Claim 7 (depends on 2) The method of claim 2, wherein the first machine learning model comprises a plurality of convolutional neural networks comprising a first convolutional neural network having a first column size and a second convolutional neural network having a second column size. ### Claim 8 (depends on 2) The method of claim 2, further comprising determining a first feature input for the first machine learning model, wherein the first feature input is generated using Bidirectional Encoder Representations from Transformers (“BERT”). ### Claim 9 (depends on 2) The method of claim 2, further comprising determining a first feature input for the first machine learning model, wherein the first feature input is generated based on textual data using natural language processing. ### Claim 10 (depends on 2) The method of claim 2, wherein determining the third output based on the weighted average of the first output and the second output comprises determining a first weight for the first output and a second weight for the second output, wherein the first weight is greater than the second weight. ### Claim 11 (depends on 10) The method of claim 10, wherein the first weight is twice the second weight. ### Claim 12 A non-transitory computer-readable medium comprising of instructions that, when executed by one or more processors, cause operations comprising: receiving a first user action during a conversational interaction with a user interface; generating a first output from a first machine learning model, wherein the first machine learning model is trained using a multi-class cross entropy loss function; generating a second output from a second machine learning model, wherein the second machine learning model is trained using a binary cross entropy loss function; determining a third output based on a weighted average of the first output and the second output; selecting a subset of dynamic conversational responses from a plurality of dynamic conversational responses based on the third output; and generating, at the user interface, the subset of dynamic conversational responses during the conversational interaction. ### Claim 13 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein determining the third output based on the weighted average of the first output and the second output comprises determining a first weight for the first output and a second weight for the second out-put, wherein the first weight is greater than the second weight. ### Claim 14 (depends on 13) The non-transitory computer-readable medium of claim 13, wherein the first weight is twice the second weight. ### Claim 15 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the first output comprises a first plurality of probabilities that summed to one, wherein each of the first plurality of probabilities corresponds to a respective user intent. ### Claim 16 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the second output comprises a second plurality of probabilities that summed do not sum to one, wherein each of the second plurality of probabilities corresponds to a respective user intent. ### Claim 17 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the instructions further cause operations comprising determining a first feature input for the first machine learning model, wherein the first feature input comprises a matrix, and wherein the first output corresponds to a prediction based on a column of the matrix and the second output corresponds to a row of the matrix. ### Claim 18 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the first machine learning model comprises training a single classifier per class, wherein samples of the class are positive samples and all other samples are negative samples. ### Claim 19 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the first machine learning model comprises a plurality of convolutional neural networks comprising a first convolutional neural network having a first column size and a second convolutional neural network having a second column size. ### Claim 20 (depends on 12) The non-transitory computer-readable medium of claim 12, wherein the instructions further cause operations comprising determining a first feature input for the first machine learning model, and wherein the first feature input is generated based on textual data using natural language processing.