Abstract: A task, such as task completed using a website, may be automated by submitting prompts to a language model and requesting that that language model provide one or more next actions to be performed to complete the task. The accuracy of the language model in providing correct actions may be improved by using one or few-shot learning where examples of completing a task are provided in a prompt to the language model. The accuracy of the language model may also be improved by breaking a task into subtasks. A prompt may be submitted to the language model to request that the language model indicate a subtask to be performed to complete the task. A prompt may then be submitted to the language model to request that the language model indicate a next action to be performed to complete the subtask.

Abstract: The extraction of key values from a conversation may be facilitated by using automation techniques. Communications of a conversation may be processed to determine a natural language intent, and the intent may be used to determine one or more target keys. The automated process may be used to obtain values for the one or more target keys. For a target key, a prompt may be presented to a user in the conversation, a response may be received from the user, and the response may be processed with a value extractor corresponding to the target key to determine a value for the target key. The values may then be presented to user in the conversation or used for other processing.

Abstract: A diffusion model may be used to generate an audio signal from text. The diffusion model may process received text and noise vectors to compute encoded audio vectors that correspond to the text. The encoded audio vectors may be decoded to generate an audio signal of a person speaking the text that may be presented to a user. The diffusion model may process a sequence of byte-encoding vectors corresponding to the text, and the use of the byte-encoding vectors may allow for the generation of higher quality audio signals. In some implementations, prompt audio of a person may also be used to generate an audio signal that resembles the speech of that person.

Abstract: A third-party service may be used to assist entities in responding to requests of users by determining a suggested resource corresponding to a received communication. The third party service may receive a request from a first entity, such as via an application programming interface request, that includes a message in a conversation. A conversation feature vector may be computed by processing the message with a first neural network. A suggested resource may be determined using the conversation feature vector. The third-party service may then return the suggested resource for use in the conversation. The third-party service may similarly be used to assist other entities in responding to requests of users.

Abstract: A message may be suggested to a user participating in a conversation using one or more neural networks where the suggested message is adapted to the preferences or communication style of the user. The suggested message may be adapted to the user with a user embedding vector that represents the preferences or communication style of the user in a vector space. To suggest a message to the user, a conversation feature vector may be computed by processing the text of the conversation with a neural network. A context score may be computed for one or more designated messages, where the context score is computed by processing the user embedding vector, the conversation feature vector, and a designated message feature vector with a neural network. A designated message may be selected as a suggested message for the user using the context scores. The suggestion may then be presented to the user.

Abstract: For any application that processes speech, improving the quality of the feature vectors may improve the quality of the speech application. The quality of feature vectors may be improved by modifying a neural network architecture for computing feature vectors to allocate computational resources where they are more effective for learning and computing the feature vectors. Contextual feature vectors may be computed from feature vectors by using a parameterized downsampling operation that decreases a vector sequence rate, processing the downsampled vectors with a neural network, and using a parameterized upsampling operation that increases a vector sequence rate. For example, parameterized downsampling may decrease a vector sequence rate by a factor of two, a neural may require fewer computational resources since it operates with a lower vector sequence rate, and parameterized upsampling may then increase the vector sequence rate by a factor of two.

Abstract: A communications session with a user may be automated using a language model. The language model may be instructed to select a next action to be performed where the next action may include transmitting a responsive communication to the user or performing an API call. The prompt used to query the language model may include one or more of the following: a representation of text of the communications session, a list of available API calls, instructions to select a next action, a representation of API calls performed, or a representation of API call responses received. The language model may be sequentially queried to continue the communications session by transmitting responsive communications or performing API calls. In some implementations, a prompt template may be used to generate the prompt and a prompt template may be selected using text of the communications session.

Abstract: The amount of future context used in a speech processing application allows for tradeoffs between performance and the delay in providing results to users. Existing speech processing applications may be trained with a specified future context size and perform poorly when used in production with a different future context size. A speech processing application trained using a stochastic future context allows a trained neural network to be used in production with different amounts of future context. During an update step in training, a future-context size may be sampled from a probability distribution, used to mask a neural network, and compute an output of the masked neural network. The output may then be used to compute a loss value and update parameters of the neural network. The trained neural network may then be used in production with different amounts of future context to provide greater flexibility for production speech processing applications.

Abstract: A statistical language model may be used to simulate one or more users of a conversation. The statistical language model may be used to train a user to participate in a particular types of conversation by simulating communications by another type of user in the conversation. The communications may be simulated by selecting a simulation context from available simulation contexts and the simulation context may correspond to a difficulty level. Upon receiving a communication from a user, a responsive simulated communication may be generated by processing the received communication and the simulation context with the statistical language model. Upon completion of the simulation, another simulation context may be selected for the next simulation.

Abstract: The present disclosure relates to a system, method, and computer program for adjusting operations of a customer service application based on metrics generated as a result of substantially real-time monitoring of entity states within the customer service application. The system receives entity events from a plurality of services associated with the customer service application. It tracks states of entities in substantially real time within the customer service application based on the events and the state machine logic for the customer service application, including identifying any entities in an anomalous state. For each non-anomalous state transition, one or more state transition measurements are calculated. The system generates metrics for the customer service application based on the state transition measurements for entities in a non-anomalous state and adjusts the operations of the customer service application in substantially real time based on the metrics.

Abstract: A system, method, and computer program are provided for text sanitization. The system builds a corpus of document vectors (including tokenizing each document, creating a vector representation based on the tokens, and building a corpus of vector representations), obtains a new document for text sanitization, tokenizes the new document, creates a new document vector based on the tokens in the new document, and accesses the corpus of document vectors. The system filters each of the tokens in the new document against a privacy threshold. The system performs a k-anonymity sanitization process such that the new document vector becomes indistinguishable from at least k other document vectors in the corpus of document vectors. The system replaces or redacts the tokens in the document flagged as unsafe. The system updates the corpus of document vectors to include the new document vector in its form prior to the filtering and k-anonymity sanitization steps.

Abstract: A third-party service may be used to assist entities in responding to requests of users by determining a suggested response to a received communication. The third party service may receive a request from a first entity, such as via an application programming interface request, that includes a message in a conversation. A conversation feature vector may be computed by processing the message with a first neural network. A suggested respond to the message may be determined by processing the conversation feature vector with a second neural network. The third-party service may then return the suggested response for use in the conversation. The third-party service may similarly be used to assist other entities in responding to requests of users.

Abstract: An intent classifier may be used to increase the efficiency of a communications system. A company may provide assistance to a first user using automated processing or by a second user manually responding to the first user. To reduce costs, the company may prefer to use automated processing for assistance where it is available. While a second user is assisting a user, a message from the first user may be processed with an intent classifier to determine that automated support is available to assist the first user, and a suggestion may be presented to the second user to transfer the first user to the automated processing, such as by presenting a button to the second user to transfer the first user to the automated processing. The second user may then transfer the first user to the automated processing and assist other users.

Abstract: A message suggestion service may use clusters of pre-approved messages to improve the quality of messages suggested to users. During a conversation, messages of the conversation may be processed with a neural network to compute a conversation encoding vector. The neural network may also be used to compute pre-approved message encoding vectors of the pre-approved messages. Distances between the conversation encoding vector and the pre-approved message encoding vectors may be used to select one or more clusters. Distances between the conversation encoding vector and the pre-approved message encoding vectors may then be used to select one or more pre-approved messages from the selected clusters. The selected pre-approved messages may then be presented as suggested messages to a user.

Abstract: A company may implement automated workflows for convenience of users or to reduce support costs. For example, allowing a user to change an address using an automated workflow may be faster or less expensive than with a human agent. In some instances, a first communications session may be started between a first user and a second user. During the first communications session, one or more communications may be processed to select an intent of the first user and a value of an information item communicated by the first user. An automated workflow may be selected to continue assisting the first user, and the first user may be transferred to a second communications session with the automated workflow. The automated workflow may be initialized with the value of the information item that was provided during the first communications session so that the first user does not need to repeat information.

Abstract: A user using a messaging application may be in conversations with multiple people and may inadvertently send a message intended for a first person to a second person. The user may be warned before making such mistakes by processing the text of an entered message and/or the text of the conversations with a mathematical model. A match score may be computed that indicates the match between the entered message and the conversation in which it was entered. Where the match score indicates a possible mistake, a warning may be presented to the user. In some implementations, a match score may be computed using a conversation encoding vector and a message encoding vector. In some implementations, a match score may be computed by processing a sequence of tokens for the conversation and the entered message that includes special token separators.

Abstract: A description of a conversation may be generated to allow a person to understand important aspects of the conversation without needing to review the conversation. The conversation description may be generated by identifying one or more events that occurred in the conversation and then generating the description using the identified events. A set of possible events may be determined in advance for a particular application. The events may be identified by using an event neural network for each event. Each event neural network may process the messages of the conversation to generate an event score that indicates a match between the conversation and the corresponding event. The event scores may then be used to select one or more events. Message scores from the event neural network of a selected event may then be used to select one or more messages of the conversation as a rationale for the selected event.

Abstract: Medical records may be analyzed to identify important items in the text of the medical record. Actionable content may be identified and may be emphasized or extracted from the medical record. Actionable content may be categorized into one or more categories. Identification may include processing using trained models that use contextual information and position information to determine sentence labels.

Abstract: A first named entity recognition (NER) system may be adapted to create a second NER system that is able to recognize a new named entity using few-shot learning. The second NER system may process support tokens that provide one or more examples of the new named entity and may process input tokens that may contain the new named entity. The second NER system may use a classifier of the first NER system to compute support token embeddings from the support tokens and input token embeddings from the input tokens. The second NER system may then recognize the new named entity in the input tokens using abstract tag transition probabilities and/or distances between the support token embeddings and the input token embeddings.

Abstract: A graphical user interface (GUI) page may be represented as GUI page encoding to facilitate processing of the GUI page in an application of GUI pages. A GUI page encoding may be computed by processing a GUI with a GUI page encoding model, and a GUI page encoding model may be trained by processing a training corpus of sequences of GUI pages. The training process may include obtaining first and second GUI pages from the training corpus, computing first and second GUI page encodings with the GUI page encoding model, computing a predicted GUI page encoding by processing the first GUI page encoding with a page predictor model, computing an error value be comparing the predicted GUI page encoding and the second GUI page encoding, and updating parameters of the GUI page encoding model by performing back propagation using the error value.