Abstract: Disclosed herein is a system and method for Natural Language Processing (NLP) of real world documents. the system and method combines various models not previously combined and overcomes the challenges of this combination. Models include an encoder-decoder model, a spatial model, and a multi-modal model. An iterative training process receives documents and generates outputs, wherein the iterative training process comprises enabling information retrieval from documents without training data.

Abstract: Apparatus, methods, and computer-readable media facilitating reduction and incremental modeling through set theory prime factorization are disclosed herein. An example method includes identifying one or more topics associated with content of a social media message. The example method also includes mapping each of the one or more topics to a respective set key to generate a set list associated with the social media message. Additionally, the example method includes recording the set list associated with the social media message in memory.

Abstract: A system and method for federated context-sensitive language models comprising a federated language model server and a plurality of edge devices. The federated language model server may comprise one or more machine learning models trained and developed centrally on the server and distribute these one or more machine learning models to edge devices wherein they may be operated locally on the edge devices. The edge devices may gather or generate context data that can be used by a speech recognition engine, and the local language models contained therein, to develop adaptive, context-sensitive, user-specific language models. Periodically, the federated language model server may select a subset of edge devices from which to receive uploaded local model parameters, that may be aggregated to perform central model updates wherein the updated model parameters may then be sent back to edge devices in order to update the local model parameters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for handing off a user conversation between computer-implemented agents. One of the methods includes receiving, by a computer-implemented agent specific to a user device, a digital representation of speech encoding an utterance, determining, by the computer-implemented agent, that the utterance specifies a requirement to establish a communication with another computer-implemented agent, and establishing, by the computer-implemented agent, a communication between the other computer-implemented agent and the user device.

Abstract: Methods and systems for performing semantic entity recognition. The method includes accessing a document stored in a memory and selecting from a general knowledge data repository, target domain information based on a specified target domain. The method also includes generating a plurality of weak annotators for the document based upon the selected target domain information and expert knowledge from a domain-specific expert knowledge data repository and applying the plurality of weak annotators to the document to generate a plurality of weak labels. The method further includes selecting at least one weak label from the plurality of weak labels as training data and training a semantic entity prediction model using the training data.

Abstract: Embodiments are directed to a companding method and system for reducing coding noise in an audio codec. A compression process reduces an original dynamic range of an initial audio signal through a compression process that divides the initial audio signal into a plurality of segments using a defined window shape, calculates a wideband gain in the frequency domain using a non-energy based average of frequency domain samples of the initial audio signal, and applies individual gain values to amplify segments of relatively low intensity and attenuate segments of relatively high intensity. The compressed audio signal is then expanded back to the substantially the original dynamic range that applies inverse gain values to amplify segments of relatively high intensity and attenuating segments of relatively low intensity. A QMF filterbank is used to analyze the initial audio signal to obtain a frequency domain representation.

Abstract: A virtual agent that utilizes an in-house natural language understanding (NLU) service and integrates a third party NLU service. The third-party NLU service is integrated with the virtual agent via a transformation script that establishes a transformation boundary through which communications are directed for adjustment and conditioning. The third party NLU service communicates with the virtual agent via an application programming interface (API). The virtual agent receives an utterance from a user via a chat session and provides the utterance to the third party NLU service. The third party NLU service may return intents, entities, and confidence, generate and return a response, and/or take actions within the cloud-based platform via the API, dependent upon the degree of integration. The virtual agent then provides a response to the user via the chat session.

Abstract: The present disclosure relates to a convolutional-neural-network structure for acquiring intent words, and a speech recognition device and method using the network. the method includes receiving input data generated from speech, performing convolution on the input data and N3 filters each having N2 channels, and acquiring a feature map having N4 pieces of data for each channel, applying max pooling to the N4 pieces of data to acquire a representative value, and acquiring a feature map having N2 pieces of data for each filter, performing concatenation on the feature maps for the respective filters, and acquiring one feature map of an N3×N2 matrix, performing convolution on the feature map of the N3×N2 matrix and a filter of a 1×N3 matrix, and acquiring a feature map of a 1×N2 matrix; and inputting the feature map of the 1×N2 matrix into an artificial neural network, and acquiring at least one intent word.

Abstract: According to embodiments of the present disclosure, a method, apparatus, device, and computer readable storage medium for voice interaction are provided. The method includes: determining a text corresponding to the voice signal based on a voice feature of a received voice signal. The method further includes: determining, based on the voice feature and the text, a matching degree between a reference voice feature of an element in the text and a target voice feature of the element. The method further includes: determining a first possibility that the voice signal is an executable command based on the text. The method further includes: determining a second possibility that the voice signal is the executable command based on the voice feature.

Abstract: An artificial intelligence apparatus for a speech interaction comprises: a microphone configured to obtain speech data corresponding to an utterance sentence of a user; a memory configured to store a first platform framework corresponding to a default platform of the artificial intelligence apparatus and at least one second platform framework; a communication unit configured to communicate with at least one of an Internet of Things (IoT) device or an external platform apparatus; and a processor configured to: obtain intention information corresponding to the speech data, determine at least one target device among the artificial intelligence apparatus, the IoT device, or the external platform apparatus based on the intention information, determine a target platform framework among the first platform framework or the second platform framework based on the intention information, generate a command corresponding to the speech data using the determined target platform framework, and perform an operation correspond

Abstract: Disclosed are a speech processing method and speech processing apparatus which execute an installed artificial intelligence (AI) algorithm and/or machine learning algorithm to perform speech processing to enable the speech processing apparatus, a user terminal, and a server to communicate with each other in a 5G communication environment. The speech processing method according to an exemplary embodiment of the present disclosure includes converting a user's spoken utterance into a user utterance text, discovering a domain, an intent, and at least one named entity of the user utterance text, determining whether the user's spoken utterance is a complete spoken utterance or an incomplete spoken utterance, generating a spoken query utterance and feeding the generated spoken query utterance back to the user when the user's spoken utterance is an incomplete spoken utterance, and constructing a complete spoken utterance by receiving a user's spoken response utterance which responds to the spoken query utterance.

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: Disclosed is an artificial intelligence (AI) apparatus for training an acoustic model, and more particularly, an AI apparatus for training an acoustic model including a shared network and a branch network connected to the shared network using speech data and phonemes corresponding to the speech data.

Abstract: A method, apparatus and computer program product are provided to evaluate natural language input to identify actions and landmarks, such as for the generation of a landmark graph. In a method, one or more word representations are encoded in context based upon one or more other word representations within a sentence of the natural language input. The method also generates an action context vector defining actions that occur during each of a plurality of time steps and generates a state context vector defining one or more landmarks associated with each of the plurality of time steps. The method predicts the state for a respective time step by defining the one or more landmarks associated with the respective time step. The method also predicts an action for the respective time step based on the action context vector and the state context vector.

Abstract: Provided is a method and a platform for compressing a pre-training language model based on knowledge distillation. According to the method, a universal knowledge distillation strategy of feature migration is firstly designed, and in the process of knowledge distillation from the teacher model to the student model, the feature mapping of each layer of the student model is approaching the teacher's features, focusing on the ability of small samples to express features in the intermediate layer of the teacher model, and guiding the student model by using these features; then, a knowledge distillation method based on self-attention cross is constructed; finally, a linear transfer strategy based on Bernoulli probability distribution is designed to gradually complete the knowledge transfer of feature mapping and self-attention distribution from teachers to students.

Abstract: In accordance with embodiments, methods and systems for a trigger to the KWS are provided. The computing device converts an audio signal into a plurality of audio frames. The computing device generates a Mel Frequency Cepstral Coefficients (MFCC) matrix. The MFCC matrix includes N columns. Each column of the N columns comprises coefficients associated with audio features corresponding to a different audio frame of the plurality of audio frames. The computing device determines that a trigger condition is satisfied based on an MFCC_0 buffer. The MFCC_0 buffer comprises a first row of the MFCC matrix. The computing device then provides the MFCC matrix to a neural network for the neural network to use the MFCC matrix to make keyword inference based on the determining that the trigger condition is satisfied.

Abstract: This disclosure relates generally to human-robot interaction (HRI) to enable a robot to execute tasks that are conveyed in a natural language. The state-of-the-art is unable to capture human intent, implicit assumptions and ambiguities present in the natural language to enable effective robotic task identification. The present disclosure provides accurate task identification using classifiers trained to understand linguistic and semantic variations. A mixed-initiative dialogue is employed to resolve ambiguities and address the dynamic nature of a typical conversation. In accordance with the present disclosure, the dialogues are minimal and directed to the goal to ensure human experience is not degraded. The method of the present disclosure is also implemented in a context sensitive manner to make the task identification effective.

Abstract: The following units are included: a word search unit searching a database for each of words included in an input word pair; a word vector extracting unit extracting a word vector corresponding to each of the searched words from a word vector database; a character vector converting unit converting characters included in the word pair into a character vector; a feature amount acquiring unit acquiring a feature amount by combining the extracted word vector of each of the words and the character vector generated by converting the characters; and an estimation unit estimating a semantic relation between the words from the acquired feature amount.

Abstract: A method is provided to implement a conversational system with artifact generation. A middleware component receives a user input and determines whether there is sufficient information in the user input and a conversation space in a context storage of the conversational system to identify user intent associated with the user input. Responsive to the middleware component determining there is not sufficient information to identify user intent, a communications handler component sends a natural language query to an external data source via a natural language query (NLQ) interface and receives a natural language response from the external data source. The middleware component updates the conversation space based on the natural language response and returns a user response based on the natural language response.

Abstract: An information processing device includes a communication controller that performs communication control for receiving transmission data transmitted from a client, transmitting the transmission data to a first service providing server that performs a first service process, receiving a first service process result from the first service providing server, transmitting data according to the first service process result to a second service providing server that performs a second service process that is different from a first service, receiving a second service process result from the second service providing server, and transmitting the second service process result to the client. The first service process result is obtained by performing the first service process on the transmission data. The second service process result is obtained by performing the second service process on the data according to the first service process result.