Google Inc.

Palo Alto Research Center Inc.

Xerox Corporation

Thorsten Brants

Jay Michael Ponte

Peng Xu

Jeffrey Adgate Dean

Francine R Chen

Franz Josef Och

Alexander Mark Franz

Jeffrey Chin

Daniel Rosart

John S Hawkins

Sha-Mayn Teh

Karel Driesen

Ignacio E Thayer

Anton Carver

Annie E Zaenen

Quoc V Le

Yonghui Wu

Ciprian Ioan Chelba

Zhifeng Chen

Ayman Farahat

Ashok Chhabedia Popat

Melvin Jose Johnson Premkumar

Michael Schuster

Sarveshwar Duddu

Ioannis Tsochantaridis

Maxim Krikun

Kuntal Loya

Minh Tue Vo Thanh

Christopher Rohrs

Paula S Newman

Shashidhar Anil Thakur

Thomas Hofmann

Hayden Shaw

Sushrut Suresh Karanjkar

David Robert Talbot

John C Blitzer

David Talbot

Jan Pfeifer

Pavel Levin

John Hawkins

Systems, methods, and apparatus for accessing distributed models in automated machine processing, including using large language models in machine translation, speech recognition and other applications.

Encoding and adaptive, scalable accessing of distributed models

Methods, systems, and apparatus, including computer programs encoded on computer storage media for performing machine learning tasks. One method includes receiving (i) a model input, and (ii) data identifying a first machine learning task to be performed on the model input to generate a first type of model output for the model input; augmenting the model input with an identifier for the first machine learning task to generate an augmented model input; and processing the augmented model input using a machine learning model, wherein the machine learning model has been trained on training data to perform a plurality of machine learning tasks including the first machine learning task, and wherein the machine learning model has been configured through training to process the augmented model input to generate a machine learning model output of the first type for the model input.

Implicit bridging of machine learning tasks

Methods, systems, and apparatus, including computer programs encoded on computer storage media for performing machine learning tasks. One method includes receiving (i) a model input, and (ii) data identifying a first machine learning task to be performed on the model input to generate a first type of model output for the model input; augmenting the model input with an identifier for the first machine learning task to generate an augmented model input; and processing the augmented model input using a machine learning model. An exemplary system applying implicit bridging for machine learning tasks, as described in this specification, trains a machine learning model to perform certain types of machine learning tasks without requiring explicit training data for the certain types of machine learning tasks to be used during training.

A server is configured to receive, from a client, a query and context information associated with a document; obtain search results, based on the query, that identify documents relevant to the query; analyze the context information to identify content; generate first scores for a hierarchy of topics, that correspond to measures of relevance of the topics to the content; select a topic that is most relevant to the context information when the topic is associated with a greatest first score; generate second scores for the search results that correspond to measures of relevance, of the search results, to the topic; select one or more of the search results as being most relevant to the topic when the search results are associated with one or more greatest second scores; generate a search result document that includes the selected search results; and send, to a client, the search result document.

Context-based filtering of search results

Systems, methods, and apparatuses, including computer program products, are provided for machine translation using information retrieval techniques. In general, in one implementation, a method is provided. The method includes providing a received input segment as a query to a search engine, the search engine searching an index of one or more collections of documents, receiving one or more candidate segments in response to the query, determining a similarity of each candidate segment to the received input segment, and for one or more candidate segments having a determined similarity that exceeds a threshold similarity, providing a translated target segment corresponding to the respective candidate segment.

Machine translation using information retrieval

Systems, methods, and computer program products for machine translation are provided. In some implementations a system is provided. The system includes a language model including a collection of n-grams from a corpus, each n-gram having a corresponding relative frequency in the corpus and an order n corresponding to a number of tokens in the n-gram, each n-gram corresponding to a backoff n-gram having an order of n&#x2212;1 and a collection of backoff scores, each backoff score associated with an n-gram, the backoff score determined as a function of a backoff factor and a relative frequency of a corresponding backoff n-gram in the corpus.

Large language models in machine translation

A translation system receives a test pair that includes a source test phrase in a first language and a target test phrase in a second language. The test pair can be evaluated by comparing its components with phrases in primary pairs. The test source phrase can be compared to a primary source phrase that is the phrase most commonly translated by the machine translation system into the test target phrase. The test target phrase can be compared to a primary target phrase that is the phrase into which the target source phrase is most often translated. If one and/or both comparisons are sufficiently dissimilar, the machine translation system can be modified by deleting the test pair, by flagging it for human review, or in other ways.

Back-translation filtering

Techniques for training and using linked event detection systems and transforming source-identified stopwords are provided. A training corpus of source identified stories and a reference language is determined. Optionally, stopwords for source-identified stories are transformed based on statistical analysis of parallel verified and un-verified transformations. Reference language and non-reference language terms are selectively included in source-pair term frequency-inverse story frequency models. Optionally, incremental source-identified term frequency-inverse story frequency models are determined. Selected terms are weighted and similarity metrics determined. Associated source-pair statistics, computed in part from a training corpus, are combined with the values of each similarity metric in the set of similarity metrics to form a similarity vector. Similarity vectors and verified link label information are used to determine a predictive model. Similarity vectors for story pairs are used with the predictive model to determine if the story-pairs are linked. Sources are arranged based on source inter-relationships into a source-hierarchy. Progressively more refined source-pair similarity statistics are also provided. New sources and associated source-pair similarity statistics are added by substituting related source-pair similarity statistics based on the source hierarchy and source characteristics. The source-pair similarity statistics are used to optionally normalize the similarity metrics.

Systems and methods for linked event detection

A semantic locator determines whether input sequences form semantically meaningful units. The semantic locator includes a coherence component that calculates a coherence of the terms in the sequence and a variation component that calculates the variation in terms that surround the sequence. A heuristics component may additionally refine results of the coherence component and the variation component. A decision component may make the determination of whether the sequence is a semantic unit based on the results of the coherence component, variation component, and heuristics component.

Semantic unit recognition

This specification describes technologies relating to providing search results. One aspect of the subject matter described in this specification can be embodied in methods that include the actions of receiving a network resource, the network resource including text content; generating a language model score for the resource including applying a language model to the text content of the resource; generating a query stuffing score for the reference, the query stuffing score being a function of term frequency in the resource content and a query index; calculating a gibberish score for the resource using the language model score and the query stuffing score; and using the calculated gibberish score to determine whether to modify a ranking score of the resource.

Identifying gibberish content in resources

Systems, methods, and apparatuses including computer program products are provided for encoding and using a language model. In one implementation, a method is provided. The method includes generating a compact language model, including receiving a collection of n-grams, each n-gram having one or more associated parameter values, determining a fingerprint for each n-gram of the collection of n-grams, identifying locations in an array for each n-gram using a plurality of hash functions, and encoding the one or more parameter values associated with each n-gram in the identified array locations as a function of corresponding array values and the fingerprint for the n-gram.

Randomized language models

Growing community of inventors

Palo Alto, CA, United States of America

Thorsten Brants