Abstract: Data is received that includes a passage of text generated in response to a prompt which comprises a plurality of sentences. Thereafter, the passage of text is tokenized into a plurality of tokens each corresponding to a different word in the passage of text. A first classification head of an adaptive fine-tuned transforms classifies each of the tokens into one of a plurality of classes. A second classification head of the adaptive fine-tuned transformer model classifies each of the sentences as either including or not including an argument. Data can then be provided which characterizes the first and second classifications. Related apparatus, systems, techniques and articles are also described.

Abstract: The present document relates to a method of layered encoding of a compressed sound representation of a sound or sound field. The compressed sound representation comprises a basic compressed sound representation comprising a plurality of components, basic side information for decoding the basic compressed sound representation to a basic reconstructed sound representation of the sound or sound field, and enhancement side information including parameters for improving the basic reconstructed sound representation.

Abstract: Examples provide enhanced machine learning model accuracy through post-hoc confidence score calibration. A machine learning (ML) system receives results generated by an ML model, the results comprising at least one confidence score and electronic documents. The ML system processes the results generated by the ML model comprising performing document understanding by extracting data points from the electronic documents. The ML system associates the confidence score with the extracted data points and calibrates a confidence score associated with the extracted data points using a post-hoc calibration solution set. The ML system implements confidence scoring recalibration comprising aligning the confidence score with prediction accuracy and adjusting the generated confidence score by the recalibration. Based on adjusting the confidence score, the ML system extracts an individual element of information from the electronic documents comprising an extracted value.

Abstract: A question answering system includes a dangerous operation determination unit configured to determine whether question-answer pair data includes a dangerous operation, and an answer sentence replacement unit configured to replace, when it is determined that the dangerous operation is included, an answer sentence including a description of the dangerous operation included in a document by a replaced answer sentence depending on a classification of the dangerous operation, to thereby create replaced question-answer pair data.

Abstract: Systems, methods and non-transitory computer readable media for generating content using a generative model without relying on selected training examples are provided. An input indicative of a desire to generate a new content using a generative model may be received. The generative model may be a result of training a machine learning model using a plurality of training examples. Each training example of the plurality of training examples may be associated with a respective content. Further, an indication of a particular subgroup of at least one but not all of the plurality of training examples may be obtained. Based on the indication, the input and the generative model may be used to generate the new content, abstaining from basing the generation of the new content on any training example included in the particular subgroup. The new content may be provided.

Abstract: In some embodiments, a method is provided for updating an editing parameter for a model for automatically suggesting revisions to text data. The method may include displaying, on a graphical user interface (GUI) of a user device, one or more interactive input elements, wherein each of the one or more input elements is associated with an editing parameter for a model for automatically suggesting revisions to text data. The method may include receiving, via the GUI, an input from a selected input element of the one or more input elements, wherein the input comprises an indication of a value for a selected editing parameter associated with the selected input element. The method may include updating the selected editing parameter for the model based on the value. The method may include using the model with the updated selected editing parameter to apply an edit operation to an obtained text-under-analysis.

Abstract: A document is obtained that include natural language text and a set of conditions are extracted from the document. In at least one embodiment, the set of conditions is extracted using a machine learning model trained on other natural language text. A validation process is performed to result in a validated set of conditions and a machine-readable data structure is generated to indicate one or more conditions of the validated set of conditions expressed within the document.

Abstract: Methods, systems, and computer-readable media for the generation of real-time recommendations using natural language processing. The method receives a request for a benefit recommendation; generates at least one tag based on input data; extracts, based on the at least one tag, at least one observation and at least one action from the input data; predicts at least one recommendation based on the extracted at least one observation and the extracted at least one action in real time; and sends the at least one predicted recommendation for display to a user device.

Abstract: The various implementations described herein include methods and devices for speaker diarization. In one aspect, a method includes obtaining an audio recording and generating an embedding signal from the audio recording. The method further includes factoring the embedding signal to obtain a basis matrix and an activation matrix, including obtaining a sparse optimization of the embedding signal by minimizing a norm corresponding to the factored embedding signal. The method also includes generating a speaker log for the audio recording based on the sparse optimization of the embedding signal.

Abstract: A computer-implemented method includes receiving a sequence of acoustic frames as input to an automatic speech recognition (ASR) model. Here, the ASR model includes a causal encoder and a decoder. The method also includes generating, by the causal encoder, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method also includes generating, by the decoder, a first probability distribution over possible speech recognition hypotheses. Here, the causal encoder includes a stack of causal encoder layers each including a Recurrent Neural Network (RNN) Attention-Performer module that applies linear attention.

Abstract: A speech recognition system comprises: an input, for receiving an input signal from at least one microphone; a first buffer, for storing the input signal; a noise reduction block, for receiving the input signal and generating a noise reduced input signal; a speech recognition engine, for receiving either the input signal output from the first buffer or the noise reduced input signal from the noise reduction block; and a selection circuit for directing either the input signal output from the first buffer or the noise reduced input signal from the noise reduction block to the speech recognition engine.

Abstract: Methods and systems are provided for customizing an action. In some implementations, voice input is received from a user and a context is determined from the voice input. Potential contextual data is identified based on the context and the voice input. A level of confidence is determined for an association of the potential contextual data and the context. An action is performed based on the voice input, the potential contextual data, and the level of confidence. The potential contextual data is used to customize the action.

Abstract: An apparatus including circuitry configured to obtain a defocus direction; process a spatial audio signal that represents an audio scene to generate a processed spatial audio signal that represents a modified audio scene based on the defocus direction, so as to control relative deemphasis in, at least in part, a portion of the spatial audio signal in the defocus direction relative to at least in part other portions of the spatial audio signal; and output the processed spatial audio signal, wherein the modified audio scene based on the defocus direction enables the deemphasis in, at least in part, the portion of the spatial audio signal in the defocus direction relative to at least in part other portions of the spatial audio signal.

Abstract: Systems, apparatuses, methods, and computer program products are disclosed for automated prototyping of a topic model. An example method includes a data manipulation engine ingesting and pre-processing source data from a set of data sources, a feature extraction engine that thereafter transforms the pre-processed data into a set of numeric representations of the pre-processed data, and an autonomous model generator that automatically generates a trained topic model using the set of numeric representations. Embodiments further enable visualization of topic model output, which permits a user to easily consume and utilize information from a topic model for any number of purposes.

Abstract: In an embodiment, a programmed computer system implemented via client-server Software as a Service (SaaS) techniques provides an interactive user interface for identifying specific portions of a digital document susceptible for review and improvement. A server computer may receive a representation of a digital document, such as an email, comprising words arranged into sentences. An embodiment may tokenize a set of all sentences comprising the sequence of sentences into a document-specific vocabulary, then compute a corresponding first and second score for each sentence of the sequence of sentences. The first score may represent a calculated probability of semantic importance of the corresponding sentence to an overall meaning of the digital document. The second score may represent a calculated likelihood that the corresponding sentence will be read by a future reader of the digital document. An embodiment may identify key sentences using the first scores and second scores.

Abstract: An analysis platform combines unsupervised and semi-supervised approaches to quickly surface and organize relevant user intentions from conversational text (e.g., from natural language inputs). An unsupervised and semi-supervised pipeline is provided that integrates the fine-tuning of high performing language models via a language models fine-tuning module, a distributed KNN-graph building method via a KNN-graph building module, and community detection techniques for mining the intentions and topics from texts via an intention mining module.

Abstract: A system for intelligent editing of legal documents. The system includes a computing device. The computing device is configured to access a plurality of legal source texts from a plurality of legal sources, generate a score for each of the plurality of legal source texts, train a natural language processing model as a function of the scored legal source texts and a first machine-learning process, receive user inputted legal text from a user device being operated by a human user to create a user legal document, analyze the user inputted legal text using the natural language processing model, suggest, as a function of the analyzing, a modification to a target text of the user inputted legal text, and generate a score for a modified user legal document. A method for intelligent editing of legal documents is also provided.

Abstract: Voice-based interaction with video content being presented by a media player application is enhanced through the use of an automated assistant capable of identifying when a spoken utterance by a user is a request to playback a specific scene in the video content. A query identified in a spoken utterance may be used to access stored scene metadata associated with video content being presented in the vicinity of the user to identify one or more locations in the video content that correspond to the query, such that a media control command may be issued to the media player application to cause the media player application to seek to a particular location in the video content that satisfies the query.

Abstract: A method of processing speech includes: providing a first set of audio data having audio features in a first bandwidth; down-sampling the first set of audio data to a second bandwidth lower than the first bandwidth; producing, by a high frequency reconstruction network (HFRN), an estimate of audio features in the first bandwidth for the first set of audio data, based on at least the down-sampled audio data; inputting, into the HFRN, a second set of audio data having audio features in the second bandwidth; producing, by the HFRN, based on a second set of audio data having audio features in the second bandwidth, an estimate of audio features in the first bandwidth for the second set of audio data; and training a speech processing system (SPS) using the estimates of audio features in the first bandwidth for the first and second sets of audio data.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for recognizing speech using a spiking neural network acoustic model implemented on a neuromorphic processor are described. In one aspect, a method includes receiving, a trained acoustic model implemented as a spiking neural network (SNN) on a neuromorphic processor of a client device, a set of feature coefficients that represent acoustic energy of input audio received from a microphone communicably coupled to the client device. The acoustic model is trained to predict speech sounds based on input feature coefficients. The acoustic model generates output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone. The neuromorphic processor updates one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data.