Sri International

Facebook, Inc.

Adobe, Inc.

Sri Internaitonal

Jing Zheng

Horacio E Franco

Ying Zhang

Jayant Kumar

Christian Fuegen

Victor Abrash

Guillaume Lample

Federico Cesari

Andreas Stolcke

Gregory Kent Myers

Necip Fazil Ayan

Leonardo Neumeyer

Michael Frandsen

Reshef Shilon

Vipul Dalal

Cregg Cowan

Kristin Precoda

Ashok Gupta

John W Butzberger

Sanat Sharma

Christopher George

Ming Liu

Manasi Deshmukh

Chirag Arora

Karthik Suresh

Ravindra Sadaphule

Andrei Stefan

Arindam Mandal

Murat Akbacak

Embodiments are disclosed for receiving a modal input including at least one of a text input or an image input. The method may include extracting an intent label from the modal input. The method may further include generating, by an intent embedding generator, an intent embedding from the intent label. The method may further include comparing the intent embedding to a plurality of candidate font embeddings to obtain one or more candidate fonts based on a similarity of the intent embedding to the plurality of candidate font embeddings in an embedding space. The method may further include identifying a recommended font based on the similarity of the intent embedding to a selected candidate font embedding of the plurality of candidate font embeddings.

Multimodal input contextual font recommendations

This disclosure describes methods, non-transitory computer readable storage media, and systems that generate a digital knowledge graph based on a plurality of tutorial content items to generate recommendations of digital resource items. Specifically, the disclosed system extracts a plurality of tasks, subject categories related to the tasks, and context signals related to an environment for the tasks from a plurality of tutorial content items for one or more digital content editing applications. The disclosed system generates a digital knowledge graph including nodes corresponding to the tasks and subject categories connected via edges based on relationships extracted from the tutorial content items. In some embodiments, the disclosed system also includes nodes corresponding to digital resource items in the digital knowledge graph or in a subgraph. The disclosed system utilizes the digital knowledge graph with context data to provide a recommendation of digital resource items for display at a client device.

Generating and utilizing a digital knowledge graph to provide contextual recommendations in digital content editing applications

Components of language processing engines, such as translation models and language models, can be customized for groups of users or based on user type values. Users can be organized into groups or assigned a value on a continuum based on factors such as interests, biographical characteristics, social media interactions, etc. In some implementations, translation engine components can be customized for groups of users by selecting the training data from content created by users in that group. In some implementations, the group identifier or continuum value can be part of the input into a general translation component allowing the translation component to take a language style of that user group into account when performing language processing tasks.

Social hash for language models

Exemplary embodiments relate to techniques to classify or detect the intent of content written in a language for which a classifier does not exist. These techniques involve building a code-switching corpus via machine translation, generating a universal embedding for words in the code-switching corpus, training a classifier on the universal embeddings to generate an embedding mapping/table; accessing new content written in a language for which a specific classifier may not exist, and mapping entries in the embedding mapping/table to the universal embeddings. Using these techniques, a classifier can be applied to the universal embedding without needing to be trained on a particular language. Exemplary embodiments may be applied to recognize similarities in two content items, make recommendations, find similar documents, perform deduplication, and perform topic tagging for stories in foreign languages.

Language-agnostic understanding

A dialog assistant embodied in a computing system can present a clarification question based on a machine-readable version of human-generated conversational natural language input. Some versions of the dialog assistant identify a clarification target in the machine-readable version, determine a clarification type relating to the clarification target, present the clarification question in a conversational natural language manner, and process a human-generated conversational natural language response to the clarification question.

Clarifying natural language input using targeted questions

The present invention relates to a method and apparatus for adding new vocabulary to interactive translation and dialog systems. In one embodiment, a method for adding a new word to a vocabulary of an interactive dialog includes receiving an input signal that includes at least one word not currently in the vocabulary, inserting the word into a dynamic component of a search graph associated with the vocabulary, and compiling the dynamic component independently of a permanent component of the search graph to produce a new sub-grammar, where the permanent component comprises a plurality of words that are permanently part of the search graph.

Method and apparatus for adding new vocabulary to interactive translation and dialogue systems

The present invention relates to a method and apparatus for adapting a language model in response to error correction. One embodiment of a method for processing an input signal including human language includes receiving the input signal and applying a statistical language model combined with a separate, corrective language model to the input signal in order to produce a processing result.

Method and apparatus for adapting a language model in response to error correction

In one embodiment, the present invention is a method and apparatus for error correction in speech recognition applications. In one embodiment, a method for recognizing user speech includes receiving a first utterance from the user, receiving a subsequent utterance from the user, and combining acoustic evidence from the first utterance with acoustic evidence from the subsequent utterance in order to recognize the first utterance. It is assumed that, if the first utterance has been incorrectly recognized on a first attempt, the user will repeat the first utterance (or at least the incorrectly recognized portion of the first utterance) in the subsequent utterance.

Method and apparatus for error correction in speech recognition applications

The present invention relates to a method and apparatus for obtaining complete speech signals for speech recognition applications. In one embodiment, the method continuously records an audio stream comprising a sequence of frames to a circular buffer. When a user command to commence or terminate speech recognition is received, the method obtains a number of frames of the audio stream occurring before or after the user command in order to identify an augmented audio signal for speech recognition processing. In further embodiments, the method analyzes the augmented audio signal in order to locate starting and ending speech endpoints that bound at least a portion of speech to be processed for recognition. At least one of the speech endpoints is located using a Hidden Markov Model.

Method and apparatus for obtaining complete speech signals for speech recognition applications

A method of speech recognition that uses hierarchical data structures that include a top level grammar and various related subgrammars, such as word, phone, and state subgrammars. A speech signal is acquired, and a probabilistic search is performed using the speech signal as an input, and using the (unexpanded) grammars and subgrammars as possible inputs. Memory is allocated to a subgrammar when a transition to that subgrammar is made during the probabilistic search. The subgrammar may then be expanded and evaluated, and the probability of a match between the speech signal and an element of the subgrammar for which memory has been allocated may be computed. Because unexpanded grammars and subgrammars take up very little memory, this method enables systems to recognize and process a larger vocabulary that would otherwise be possible. This method also permits grammars and subgrammars to be added, deleted, or selected by a remote computer while the speech recognition system is operating, allowing speech recognition systems to have a nearly unlimited vocabulary.

Method of dynamically altering grammars in a memory efficient speech recognition system

Growing community of inventors

San Jose, CA, United States of America

Jing Zheng