Abstract: In some applications, it may be desired to process a message to determine an intent of the message, where the intent indicates the meaning of the message. An intent classifier may be used to determine the meaning of a message by processing the message to compute a message embedding vector that represents the message in a vector space. Each possible intent may be represented by a prototype vector, and the intent of the message may be determined by comparing the message embedding to one or more prototype vectors, such as by selecting an intent whose prototype vector is closest to the message embedding. An intent classifier may be used, for example, (i) to implement an automated communications system with states where each state is associated with a subset of the possible intents or (ii) for processing usage data of a communications system to update the intents of the communications system.

Abstract: A first application programming interface (API) with a first schema may be adapted to work with a second API with a second schema using mappings of schema properties and a directed graph. An API call specification of the first API may receive first API schema properties as input and provide first API schema properties as outputs. The first API schema properties may be mapped to corresponding second API schema properties, such as using semantic representations of the schema properties. An implementation of an API call for the first API may be created by using the schema mappings and selecting a path from a directed graph corresponding to the second API, where the path includes a node corresponding to an API call of the second API. Computer code may be generated using nodes of the path, and the computer code may be stored for later use.

Abstract: The present disclosure describes a system, method, and computer program for predicting sentiment labels for audio speech utterances using an audio speech sentiment classifier pretrained with pseudo sentiment labels. A speech sentiment classifier for audio speech (“a speech sentiment classifier”) is pretrained in an unsupervised manner by leveraging a pseudo labeler previously trained to predict sentiments for text. Specifically, a text-trained pseudo labeler is used to autogenerate pseudo sentiment labels for the audio speech utterances using transcriptions of the utterances, and the speech sentiment classifier is trained to predict the pseudo sentiment labels given corresponding embeddings of the audio speech utterances. The speech sentiment classifier is then subsequently fine tuned using a sentiment-annotated dataset of audio speech utterances, which may be significantly smaller than the unannotated dataset used in the unsupervised pretraining phase.

Abstract: The present disclosure sets forth a marginal distillation approach to obtaining a unified name-entity recognition (NER) student model from a plurality of pre-trained teacher NER models with different tag sets. Knowledge from the teacher models is distilled into a student model without requiring access to the annotated training data used to train the teacher models. In particular, the system receives a tag hierarchy that combines the different teacher tag sets. The teacher models and the student model are applied to a set of input data sequence to obtain tag predictions for each of the models. A distillation loss is computed between the student and each of the teacher models. If teacher's predictions are less fine-grained than the student's with respect to a node in the tag hierarchy, the student's more fine-grained predictions for the node are marginalized in computing the distillation loss.

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 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 presented to the user.

Abstract: When processing a request containing personal information, personal information may be needed to respond to the request, but it may be desired to limit storage of personal information for privacy reasons. To accomplish both objectives, text of a message may be processed using multiple levels of redaction. A first level of redaction may replace digits of the text with a token so that sensitive numbers, such as credit card numbers or social security numbers are no longer present. A second level of redaction may replace one or more words of the text with a label indicating the text that was present, such as by replacing an address with a label indicating that an address was present or by replacing a credit card number with a label indicating that a credit card number was present.

Abstract: Machine learning models may be used during a communications session to process natural language communications and perform actions relating to the communications session. For example, a machine learning model may be used to provide an automated response to a user, to suggest a completion of text being entered by a user, or to provide information about a relevant resource. Machine learning models may rely on machine learning model data that is updated during a communications session as communications are processed by the machine learning model. To improve the performance of a machine learning model when a user leaves a first communications session and enters a second communications session, the machine learning model data may be stored during a first communications session and then retrieved during the second communications session to initialize a machine learning model for the second communications session.

Abstract: Software for a computer system may be automatically generated to reduce costs. Software may be automatically generated using a set of software components where each component may have one or more input properties and one or more output properties. A property may correspond to a type of data (e.g., a customer ID) used by the computer system. A graph may be created from the components where the graph includes component nodes for the components and property nodes for the input and output properties. To automatically generate software for a task, a task specification may be received that includes a task input property and a task output property. A path on the graph may be determined from the task input property and the task output property, and software for accomplishing the task may be generated using components on the path. The software may then be executed or stored for later execution.

Abstract: The present disclosure describes a system, method, and computer program for matching an input file to one of a plurality of datastore files and displaying the rationale for the match. Neural networks are trained to create vector representations of objects in the input file and the datastore files. The cost of each possible pairing of vector representations between the input file and a datastore file is computed, and an optimal transport algorithm is used to identify the vector pairings that result in the lowest total cost of alignment. The datastore file with the lowest total cost of alignment to the input file is identified as the best matching file. The alignment results are used to display the rationale for the match. To constrain the alignment results of the optimal transport algorithm, one or more dummy points and, in certain embodiments, duplicate points are added to one or both of the vector sets to achieve alignments with the desired sparsity patterns.

Abstract: Text of an incomplete message entered by a user may be processed using a neural network to suggest messages similar to the message the user is in the process of entering. Word embeddings may be obtained for the words of the text that represent the words in a first vector space. The word embeddings may then be processed by the neural network to compute an input message feature vector that represents the incomplete message in a second vector space. The input message feature vector may be used to select a first designated message as a suggestion from a set of designated messages, and the first designated message may be selected using a similarity score computed from the input message feature vector and a first designated message feature vector corresponding to the first designated message. The first designated message may then be presented as a suggestion to the user.

Abstract: The present disclosure relates to a system, method, and computer program for creating a complete transcription of an audio recording from separately transcribed redacted and unredacted words. The system receives an original audio recording and redacts a plurality of words from the original audio recording to obtain a modified audio recording. The modified audio recording is outputted to a first transcription service. Audio clips of the redacted words from the original audio recording are extracted using word-level timestamps for the redacted words. The extracted audio clips are outputted to a second transcription service. The system receives a transcription of the modified audio recording from the first transcription service and transcriptions of the extracted audio clips from the second transcription service.

Abstract: The amount of time required to train a neural network may be decreased by modifying the neural network to allow for greater parallelization of computations. The computations for cells of the neural network may be modified so that the matrix-vector multiplications of the cell do not depend on a previous cell and thus allowing the matrix-vector computations to be performed outside of the cells. Because the matrix-vector multiplications can be performed outside of the cells, they can be performed in parallel to decrease the computation time required for processing a sequence of training vectors with the neural network. The trained neural network may be applied to a wide variety of applications, such as performing speech recognition, determining a sentiment of text, determining a subject matter of text, answering a question in text, or translating text to another language.

Abstract: Machine learning-based approaches are used to detect trends of behavior and anomalous events from customer support messages between customers and customer support agents or other appropriate resources in an electronic environment. For example, for a plurality of time periods, a prediction model can be trained. The prediction models can be trained on messages that correspond to each prediction models' period of time. The prediction models can process messages to determine a score (e.g., a representative confidence score) for the time period a prediction model is associated with. For a selected time period, a model (e.g., a trend detection model) can be applied to the scores for time periods before the selected time period to determine whether the score for the selected time period is associated with an anomalous event. Thereafter, an alert can be presented with, for example, the messages that triggered the alert, among other such information.

Abstract: A plurality of conversations may be processed to obtain one or more representative conversations to allow a better understanding of the plurality of conversations. A representative conversation may be determined by representing each conversation as a sequence of states where a state may represent messages with similar meanings. Distances may be computed between pairs of conversations, and the conversations may be clustered using the distances. To obtain a representative conversation for a cluster of conversations, a representative sequence of states may be obtained for the cluster and a representative message may be obtained for each state of the sequence of states. The representative conversation may then be presented to a user.

Abstract: During a conversation between a customer and a customer support representative, suggestions may be presented to the customer support representative to upsell a product to the customer. Information about the customer and/or information about the conversation may be processed by a computer to determine when to suggest the upsell to the customer support representative and the one or more products to be upsold. The determination may be performed by computing features from the information about the customer and the information about the conversation, and processing the features with one or more classifiers.

Abstract: A neural network may be used to remove personal information from text (such as names, addresses, credit card numbers, or social security numbers), and replace the personal information with a label indicating the type or class of the removed information. The neural network may comprise multiple layers that compute a context vector for words of the text, compute label scores for words of the text using the context vectors, and select a label for each word using the label scores. Words corresponding to certain labels may be replaced with a label, such as replacing the digits of a credit card number with a label <cc_number>. The redacted text may then be presented to a person or stored for later processing.

Abstract: Users may request assistance or information from a limited number of resources, such as submitting a user request by speaking or entering text. A user request from among the pending user requests may be selected using a selection model. A selection model may process features relating to each of the pending user requests and generate a probability distribution for the pending user requests. A user request may then be selected using the probability distribution, such as by making a random selection. The selection model may be updated over multiple time periods by computing reward scores for the selection decisions made by the selection model and using the reward scores to update the parameters of the selection model.

Abstract: In some applications, it may be desired to process a message to determine an intent of the message, where the intent indicates the meaning of the message. An intent classifier may be used to determine the meaning of a message by processing the message to compute a message embedding vector that represents the message in a vector space. Each possible intent may be represented by a prototype vector, and the intent of the message may be determined by comparing the message embedding to one or more prototype vectors, such as by selecting an intent whose prototype vector is closest to the message embedding. An intent classifier may be used, for example, (i) to implement an automated communications system with states where each state is associated with a subset of the possible intents or (ii) for processing usage data of a communications system to update the intents of the communications system.

Abstract: A third-party service may be used to assist entities in responding to requests of users. A third-party service may receive, directly or indirectly, a request of a first user for assistance from a first entity. The third-party service may request information about the first user by sending a request to a computer of the first entity. The third-party service may use the request of the first user and the information about the first user to automatically generate a response to the request of the first user. The third-party service may then transmit, directly or indirectly, the response to the first user.

Abstract: A third-party company may assist other companies in providing customer support to their customers. The third-party company may provide software to a computer of a customer service representative to present a user interface to assist the customer service representative in responding to customer requests. Third-party company may also send update data to the computer of the customer service representative to cause a portion of the user interface to be updated, where the update data is determined using an intent of a message received from a customer. A message received from the customer may be processed to determine the intent of the message, a template may be obtained using the intent, and the update data may be generated by rendering the selected template. The update data may then be transmitted to the computer of the customer service representative to cause a portion of the user interface to be updated.