Abstract:
A way of generating a watermark for a structured result, such as a search result or a machine translation. A hash function is used to generate a bit sequence for each of a plurality of structured results. A ranking score is generated for each resulting bit sequence. The ranking score can be based on the detectability of the bit sequence compared to a randomly-generated bit sequence and the quality of each of the structured results. A structured result is selected as the watermarked structured result based upon the ranking score.
Abstract:
Methods, systems, and apparatus, including computer program products, for language translation are disclosed. In one aspect, a method includes accessing a translation hypergraph that represents a plurality of candidate translations, the translation hypergraph including a plurality of paths including nodes connected by edges; calculating first posterior probabilities for each edge in the translation hypergraph; calculating second posterior probabilities for each n-gram represented in the translation hypergraph based on the first posterior probabilities; and performing decoding on the translation hypergraph using the second posterior probabilities to convert a sample text from a first language to a second language.
Abstract:
A system with a nonstatistical translation component integrated with a statistical translation component engine. The same corpus may be used for training the statistical engine and also for determining when to use the statistical engine and when to use the translation component. This training may use probabilistic techniques. Both the statistical engine and the translation components may be capable of translating the same information, however the system determines which component to use based on the training. Retraining can be carried out to add additional components, or when after additional translator training.
Abstract:
Systems, methods, and apparatuses including computer program products for machine learning. A method is provided that includes determining model parameters for a plurality of feature functions for a linear machine learning model, ranking the plurality of feature functions according to a quality criterion, and selecting, using the ranking, a group of feature functions from the plurality of feature functions to update with the determined model parameters.
Abstract:
Systems, methods, and apparatuses including computer program products are provided for training machine learning systems. In some implementations, a method is provided. The method includes receiving a collection of phrases, normalizing a plurality of phrases of the collection of phrases, the normalizing being based at least in part on lexicographic normalizing rules, and generating a normalized phrase table including a plurality of key-value pairs, each key value pair includes a key corresponding to a normalized phrase and a value corresponding to one or more un-normalized phrases associated with the normalized key, each un-normalized phrase having one or more parameters.
Abstract:
Systems, methods, and apparatuses including computer program products for machine learning. A method is provided that includes determining model parameters for a plurality of feature functions for a linear machine learning model, ranking the plurality of feature functions according to a quality criterion, and selecting, using the ranking, a group of feature functions from the plurality of feature functions to update with the determined model parameters.
Abstract:
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.
Abstract:
Methods, systems, and apparatus, including computer program products, for language translation are disclosed. In one implementation, a method is provided. The method includes determining, for a plurality of feature functions in a translation lattice, a corresponding plurality of error surfaces for each of one or more candidate translations represented in the translation lattice; adjusting weights for the feature functions by traversing a combination of the plurality of error surfaces for phrases in a training set; selecting weighting values that minimize error counts for the traversed combination; and applying the selected weighting values to convert a sample of text from a first language to a second language.
Abstract:
A system with a nonstatistical translation component integrated with a statistical translation component engine. The same corpus may be used for training the statistical engine and also for determining when to use the statistical engine and when to use the translation component. This training may use probabilistic techniques. Both the statistical engine and the translation components may be capable of translating the same information, however the system determines which component to use based on the training. Retraining can be carried out to add additional components, or when after additional translator training.
Abstract:
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing optical character recognition. In one aspect, a method includes receiving a text image I. A set of feature functions are evaluated for a log linear model to determine respective feature values for the text image I, wherein each feature function hi maps the text image I to a feature value, and wherein each feature function hi is associated with a respective feature weight λi. A transcription {circumflex over (T)} is determined that minimizes a cost of the log linear model.