Abstract: Systems and methods are disclosed for processing forms to automatically adjudicate religious exemptions. The method includes receiving an input from a user to data fields of forms associated with a religious exemption request, wherein the input is in a first data format and includes location information, religious information, employment information, or demographic information associated with the user. Exemption-relevant features are determined from the input. A data object including the exemption-relevant features is generated. The exemption-relevant features are transformed into corresponding embeddings in a second data format, wherein the embeddings represent semantic relations between the exemption-relevant features. The authenticity of the data object is determined based on the embeddings using a classification model that has been trained using a plurality of embeddings representative of a plurality of exemption-relevant features.

Abstract: An electronic device and a method of controlling an electronic device are provided. The method includes obtaining an image comprising a text through a camera, identifying an input text, among texts included in the image, to be translated, obtaining a first vector corresponding to the input text by inputting the identified input text to an encoder of a translation model, identifying whether additional information is necessary to translate the input text by inputting the first vector to a first artificial intelligence model trained to translate the input text, based on identification that the additional information is necessary, identifying additional information among the at least one context information by inputting the first vector and at least one context information obtained from the image, and obtaining an output text corresponding to the input text by inputting the first vector and the identified additional information to a decoder of the translation model.

Abstract: Techniques are described herein for selectively processing a user's utterances captured prior to and after an event that invokes an automated assistant to determine the user's intent and/or any parameters required for resolving the user's intent. In various implementations, respective measures of fitness for triggering responsive action by the automated assistant may be determined for pre-event and a post-event input streams. Based on the respective measures of fitness, one or both of the pre-event input stream or post-event input stream may be selected and used to cause the automated assistant to perform one or more responsive actions.

Abstract: In some aspects, the techniques described herein relate to a method including: providing one or more curation content machine learning models, wherein the one or more curation content machine learning models are configured to classify a textual review into a plurality of classifications, and wherein the plurality of classifications have been assigned respective relative weights; receiving, at a collaboration service, a user review of a travel objective from a user of the collaboration service; providing the review as an input feature dataset to the one or more curation content machine learning models; receiving, as output of the one or more curation content machine learning models, a plurality of classifications of the user review; and aggregating, by the collaboration service, a relative value score based on the assigned relative weight of the plurality of classifications of the user review.

Abstract: An embodiment provides a method including obtaining language input data and providing the language input data to a first generative language model and a second generative language model. A first response from the first generative language model and a second response from a second generative language model are obtained. An indication is provided of a difference between the first response from the first generative language model and the second response from the second generative language model.

Abstract: A vector acquisition method includes: inputting, into a learned model, at least one piece of text including at least two of a plurality of words obtained by dividing a compound, the compound being a word divisible into the plurality of words; outputting, from the learned model, an adjusted vector corresponding to at least one of the words obtained by dividing the compound in the input piece of text; and acquiring a compound vector corresponding to the compound using the adjusted vector output from the learned model. Classification accuracy of vectors corresponding to words can thereby be enhanced.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, that translate text depicted in images from a source language into a target language. Methods can include obtaining a first image that depicts first text written in a source language. The first image is input into an image translation model, which includes a feature extractor and a decoder. The feature extractor accepts the first image as input and in response, generates a first set of image features that are a description of a portion of the first image in which the text is depicted is obtained. The first set of image features are input into a decoder. In response to the input first set of image features, the decoder outputs a second text that is a predicted translation of text in the source language that is represented by the first set of image features.

Abstract: A voice conversion device is provided with a linguistic information extraction unit that extracts linguistic information corresponding to utterance content from a conversion source voice signal, an appearance feature extraction unit that extracts appearance features expressing features related to the look of a person's face from a captured image of the person, and a converted voice generation unit that generates a converted voice on a basis of the linguistic information and the appearance features.

Abstract: There is a need for more effective and efficient predictive natural language summarization. This need is addressed by applying hybrid extractive and abstractive summarization techniques in a unique processing pipeline to generate a cohesive and comprehensive summary of a multi-party interaction.

Abstract: An electronic device including a speaker; a microphone; a memory configured to store a voice recognition model; and a processor configured to: identify, while a first sound signal is input through the microphone, whether a wake-up word is included in a second sound signal that is output through the speaker by inputting the second sound signal into a wake-up word detection model, and identify, based on identifying that the wake-up word is not included in the second sound signal, whether the wake-up word is included in the first sound signal by inputting the first sound signal into the wake-up word detection model.

Abstract: Techniques for processing speculative data using data history information are described. A system may receive, at a first component, first data and second data for use in a first operation. The component may use metadata associated with each of the first data and the second data to determine that the first data and the second data correspond to different versions of data of the same data type. Based on such a determination, the component may cease processing corresponding to the first operation with respect to the first data and the second data.

Abstract: A method for generating a pre-trained language model, includes: obtaining sample files; obtaining typography structure information and text information of the sample files by parsing the sample files; obtaining a plurality of task models of a pre-trained language model; obtaining a trained pre-trained language model by jointly training the pre-trained language model and the plurality of task models according to the typography structure information and the text information; and generating a target pre-trained language model by fine-tuning the trained pre-trained language model according to the typography structure information and the text information.

Abstract: Techniques for personalized batch and streaming speech-to-text transcription of audio reduce the error rate of automatic speech recognition (ASR) systems in transcribing rare and out-of-vocabulary words. The techniques achieve personalization of connectionist temporal classification (CT) models by using adaptive boosting to perform biasing at the level of sub-words. In addition to boosting, the techniques encompass a phone alignment network to bias sub-word predictions towards rare long-tail words and out-of-vocabulary words. A technical benefit of the techniques is that the accuracy of speech-to-text transcription of rare and out-of-vocabulary words in a custom vocabulary by automatic speech recognition (ASR) system can be improved without having to train the ASR system on the custom vocabulary. Instead, the techniques allow the same ASR system trained on a base vocabulary to realize the accuracy improvements for different custom vocabularies spanning different domains.

Abstract: A conversation support system is provided at an utterance place where utterance is delivered to a plurality of persons. The persons are each an utterer having a possibility of uttering and/or a performer having a possibility of marking. The conversation support system includes a hardware processor and a marking motion catcher. The hardware processor obtains voice data of an utterance made by an utterer and received by a voice receiver, and manages the voice data on a voice timeline. The marking motion catcher catches a marking motion by which a marker is given to the utterance. The hardware processor manages the marking motion on a marking timeline, and links the marking motion with the utterance on a same timeline.

Abstract: A system and method for generating a multi-label classifier of textual data is presented. The method includes training a plurality of single-label classifiers, wherein each of the plurality of single-label classifiers is trained to classify textual data to a single predefined revenue-based label; and training the multi-label classifier using labeled data output by the plurality of single-label classifiers, wherein the multi-label classifier is trained to classify textual data to a vector including revenue-based labels and their respective classification scores.

Abstract: The present disclosure presents a generative model configured to receive input regarding an item in two different modalities, such as text data and non-text data (including, for example, image or audio data), in order to generate output regarding the item that is determined based on a combination of both modalities' input. Specific relative positional and token type embeddings may be employed in an encoder portion of an encoder-decoder arrangement. An associated decoder may be trained to generate new text corresponding to diverse tasks based on the encoded representation of the two inputs as generated within the encoder. For example, the decoder may be utilized to generate attributes regarding the input item, auto-complete or auto-correct a title or description of the item, among other uses.

Abstract: Systems and processes for operating a digital assistant are provided. In one example, a method includes receiving a first speech input from a user. The method further includes identifying context information and determining a user intent based on the first speech input and the context information. The method further includes determining whether the user intent is to perform a task using a searching process or an object managing process. The searching process is configured to search data, and the object managing process is configured to manage objects. The method further includes, in accordance with a determination the user intent is to perform the task using the searching process, performing the task using the searching process; and in accordance with the determination that the user intent is to perform the task using the object managing process, performing the task using the object managing process.

Abstract: A method of document processing is provided. An implementation solution is: obtaining target text information and target layout information of a target document, the target text information includes target text included in the target document and character position information of the target text, and the target layout information is used to characterize the region where text in the target document is located; fusing the target text information and the target layout information to obtain first multimodal information of the target document; and inputting the first multimodal information into an intelligent document comprehension model, and obtaining at least one target word in the target document and at least one feature vector corresponding to the at least one target word output by the intelligent document comprehension model, each target word is related to semantics of the target document.

Abstract: Devices and techniques are generally described for inference reduction in natural language processing using semantic similarity-based caching. In various examples, first automatic speech recognition (ASR) data representing a first natural language input may be determined. A cache may be searched using the first ASR data. A first skill associated with the first ASR data may be determined from the cache. In some examples, first intent data representing a semantic interpretation of the first natural language input data may be determined by using a first natural language process associated with the first skill.

Abstract: Provided is an encoding method according to various example embodiments and an encoder performing the method. The encoding method includes outputting a linear prediction (LP) coefficients bitstream and a residual signal by performing a linear prediction analysis on an input signal, outputting a first latent signal obtained by encoding a periodic component of the residual signal, using a first neural network module, outputting a first bitstream obtained by quantizing the first latent signal, using a quantization module, outputting a second latent signal obtained by encoding an aperiodic component of the residual signal, using the first neural network module, and outputting a second bitstream obtained by quantizing the second latent signal, using the quantization module, wherein the aperiodic component of the residual signal is calculated based on a periodic component of the residual signal decoded from the quantized first latent signal output by de-quantizing the first bitstream.