Amazon Technologies, Inc.

Chengwei Su

Wael Hamza

Rahul Gupta

Spyridon Matsoukas

Sankaranarayanan Ananthakrishnan

Ahmet Emre Barut

Shirin Saleem

Kelly James Vanee

Konstantine Arkoudas

Liwei Cai

Subendhu Rongali

Beiye Liu

Weitong Ruan

Melanie C B Gens

Melanie Chie Bomke Gens

Chungnam Chan

Fan Yang

Kavya Ravikumar

Mallory McManamon

Yugang Li

Matthew Cavell Johnson

Prashan Wanigasekara

John Will Crimmins

Kechin Qin

Spurthideepika Sandiri

Luca Soldaini

John Pelak

Techniques for performing an action with respect to displayed content are described. A natural language interpretation corresponding to a received spoken user input may be determined. Prior to receiving the spoken user input, content may be displayed to the user from which the spoken user input was received. The natural language interpretation may represent a request to perform an action with respect to a portion of the content currently being displayed. Content identifiers corresponding to content being displayed, may be determined, and embedding data representing at least one feature of the content may be determined using the content identifiers. The natural language interpretation and the embedding data may be processed to determine that the spoken user input relates to a first portion of the displayed content instead of a second portion of the displayed content. Based on the determination, an action responsive to the spoken user input may be performed.

Natural language interactions using visual understanding

Devices and techniques are generally described for selection of objects in image data using natural language input. In various examples, first image data representing at least a first object and first natural language data may be received. In some examples, first embedding data representing the first natural language data may be generated. Second embedding data representing the first image data may be generated. Relative location data indicating a location of the first object in the first image data relative to at least one other object may be generated. The first embedding data, the second embedding data, and the relative location data may be input into a multi-modal transformer model. The multi-modal transformer model may determine that the first natural language data relates to the first object.

Natural language selection of objects in image data

Techniques for performing spoken language understanding (SLU) processing are described. An SLU component may include an audio encoder configured to perform an audio-to-text processing task and an audio-to-NLU processing task. The SLU component may also include a joint decoder configured to perform the audio-to-text processing task, the audio-to-NLU processing task and a text-to-NLU processing task. Input audio data, representing a spoken input, is processed by the audio encoder and the joint decoder to determine NLU data corresponding to the spoken input.

Speech processing

Techniques for processing a user input are described. Text data representing a user input is processed with respect to at least one finite state transducer (FST) to generate at least one FST hypothesis. Context information may be required to traverse one or more paths of the at least one FST. The text data is also processed using at least one statistical model (e.g., perform intent classification, named entity recognition, and/or domain classification processing) to generate at least one statistical model hypothesis. The at least one FST hypothesis and the at least one statistical model hypothesis are input to a reranker that determines a most likely interpretation of the user input.

Natural language understanding processing

A natural language understanding system that can determine an overall score for a natural language hypothesis using hypothesis-specific component scores from different aspects of NLU processing as well as context data describing the context surrounding the utterance corresponding to the natural language hypotheses. The individual component scores may be input into a feature vector at a location corresponding to a type of a device captured by the utterance. Other locations in the feature vector corresponding to other device types may be populated with zero values. The feature vector may also be populated with other values represent other context data. The feature vector may then be multiplied by a weight vector comprising trained weights corresponding to the feature vector positions to determine a new overall score for each hypothesis, where the overall score incorporates the impact of the context data. Natural language hypotheses can be ranked using their respective new overall scores.

Contextual natural language processing

A natural language processing system that can determine an overall score for a natural language hypothesis using hypothesis-specific component scores from different aspects of NLU processing. The individual component scores may be weighted by weights trained to optimize the overall scores relative to each other. Each domain of the system may be configured with a separate component that determines the overall score with respect to the domain. Natural language hypotheses can be ranked using the overall score either within a specific domain or on a cross-domain basis.

Scoring of natural language processing hypotheses

Growing community of inventors

Belmont, MA, United States of America

Chengwei Su