Abstract: Reinforcement Learning from Human Feedback (RLHF) and derivative techniques like Direct Preference Optimization (DPO) are task-alignment algorithms used to repurpose general, foundational models for specific tasks. Applying task-alignment to neural machine translation (NMT) addresses an existing task-data mismatch in NMT, leading to improvements across all languages of a multilingual model, even when task-alignment is only applied to a subset of those languages. In an embodiment, such improvements are provided by introducing Direct Quality Optimization (DQO), a variant of DPO leveraging a pre-trained translation quality estimation model as a proxy for human preferences. The improvements can be verified with both automatic metrics and human evaluation.

Abstract: A computer-implemented method is programmed to support efficient and rapid generation of machine translation suggestions on client devices. Network latency is substantially reduced or eliminated by separating certain aspects of the translation workload across multiple classes of tasks, including final neural network output, between a client device and server device. The client device and server device may be connected such that a decoder portion of a machine translation system may be downloaded onto the client device, along with an initial translation suggestion and encoder outputs associated with a document, which document is in a source language to be translated into a target language. The initial translation suggestion may be replaced by an updated machine translation suggestion as a user inputs text in the target language called a prefix. This updated machine translation is generated on the client-side decoder using the previously-downloaded encoder outputs as input and the prefix as constraint.

Abstract: A computer-implemented method is programmed to support efficient and rapid generation of machine translation suggestions on client devices. Network latency is substantially reduced or eliminated by separating certain aspects of the translation workload across multiple classes of tasks, including final neural network output, between a client device and server device. The client device and server device may be connected such that a decoder portion of a machine translation system may be downloaded onto the client device, along with an initial translation suggestion and encoder outputs associated with a document, which document is in a source language to be translated into a target language. The initial translation suggestion may be replaced by an updated machine translation suggestion as a user inputs text in the target language called a prefix. This updated machine translation is generated on the client-side decoder using the previously-downloaded encoder outputs as input and the prefix as constraint.

Abstract: An apparatus has a network interface circuit to receive a source sentence from a network connected client device. A processor is connected to the network interface circuit. A memory is connected to the processor. The memory stores parameters of a multilingual neural review system and instructions executed by the processor to operate the multilingual neural review system trained on a corpus of source sentences, draft target sentences and corrected target sentences. The multilingual neural review system produces a corrected target sentence from a draft target sentence representing a proposed translation of the source sentence. The draft target sentence and the corrected target sentence are supplied to the network connected client device. Approval for the corrected target sentence is received from the network connected client device. Parameters of the multilingual neural review system are updated based upon the approval for the corrected target sentence.

Abstract: A computer-implemented method is programmed to support efficient and rapid generation of machine translation suggestions on client devices. Network latency is substantially reduced or eliminated by separating certain aspects of the translation workload across multiple classes of tasks, including final neural network output, between a client device and server device. The client device and server device may be connected such that a decoder portion of a machine translation system may be downloaded onto the client device, along with an initial translation suggestion and encoder outputs associated with a document, which document is in a source language to be translated into a target language. The initial translation suggestion may be replaced by an updated machine translation suggestion as a user inputs text in the target language called a prefix. This updated machine translation is generated on the client-side decoder using the previously-downloaded encoder outputs as input and the prefix as constraint.

Abstract: An apparatus has a network interface circuit to receive a source sentence from a network connected client device. A processor is connected to the network interface circuit. A memory is connected to the processor. The memory stores parameters of a multilingual neural review system and instructions executed by the processor to operate the multilingual neural review system trained on a corpus of source sentences, draft target sentences and corrected target sentences. The multilingual neural review system produces a corrected target sentence from a draft target sentence representing a proposed translation of the source sentence. The draft target sentence and the corrected target sentence are supplied to the network connected client device. Approval for the corrected target sentence is received from the network connected client device. Parameters of the multilingual neural review system are updated based upon the approval for the corrected target sentence.

Abstract: An apparatus has a network interface circuit to receive a source sentence from a network connected client device. A processor is connected to the network interface circuit. A memory is connected to the processor. The memory stores parameters of a multilingual neural review system and instructions executed by the processor to operate the multilingual neural review system trained on a corpus of source sentences, draft target sentences and corrected target sentences. The multilingual neural review system produces a corrected target sentence from a draft target sentence representing a proposed translation of the source sentence. The draft target sentence and the corrected target sentence are supplied to the network connected client device. Approval for the corrected target sentence is received from the network connected client device. Parameters of the multilingual neural review system are updated based upon the approval for the corrected target sentence.

Abstract: An apparatus has a network interface circuit to receive a source sentence from a network connected client device. A processor is connected to the network interface circuit. A memory is connected to the processor. The memory stores translation data and instructions executed by the processor. The instructions executed by the processor operate a neural machine translation system. A translation hypothesis is formed from a prefix of a target sentence comprising an initial sequence of target words supplied by a user through an interface. The hypothesis is generated by the neural machine translation system that performs a constrained prefix decoding that repeatedly predicts a next word from previous target words. A suffix of the target sentence comprising a final sequence of words corresponding to a final sequence of words in the source sentence is formed using a beam search that constrains translation to match the prefix.

Abstract: A computer-implemented technique can include receiving, at a server including one or more processors, a source word in a source language. The technique can include determining, at the server, one or more potential translations for the source word in a target language different than the source language. The technique can include determining, at the server, one or more synonyms for each of the one or more potential translations to obtain a plurality of potential translations. The technique can include determining, at the server, one or more translation clusters using the plurality of potential translations and a clustering algorithm. Each translation cluster can contain all of the plurality of potential translations that have a similar denotation and each of the plurality of translations that have a similar denotation can be included in a specific translation cluster. The technique can also include outputting, at the server, the one or more translation clusters.

Abstract: Systems and methods are provided for translating a phrase block and presenting meanings and translations of the phrase block. Consistent with certain embodiments, computer-implemented systems and methods are provided for identifying parts of speech of the phrase block, determining definitions of the phrase block in the source language corresponding to the parts of speech, and determining usage examples corresponding to the definitions. Additionally, systems and methods are provided for translating the phrase block from the source language to translated text and for ranking the parts of speech based on usage. Consistent with certain embodiments, computer-implemented systems and methods are also provided for determining usage based on a user's location, based on a corpus of documents, and/or based on content stored by the user.

Abstract: Computer-implemented techniques include receiving a phrase in a first language and obtaining a corpus comprising a plurality of phrases in the first language and word reordering information for the plurality of phrases, the word reordering information indicating a correct word order for each phrase in a second language. Word-to-word correspondences between each of the phrases in the first language and the corresponding correct word order for the phrase in the second language are identified and at least one tree that allows for the identified word-to-word correspondences is generated. Based upon the at least one tree, a statistical model for reordering from a word order that is correct for the first language to a word order that is correct for the second language is created. Based upon the statistical model, a reordered phrase from the received phrase is generated, the reordered phrase having a correct word order for the second language.

Abstract: A computer-implemented technique can include receiving, at a server including one or more processors, a source word in a source language. The technique can include determining, at the server, one or more potential translations for the source word in a target language different than the source language. The technique can include determining, at the server, one or more synonyms for each of the one or more potential translations to obtain a plurality of potential translations. The technique can include determining, at the server, one or more translation clusters using the plurality of potential translations and a clustering algorithm. Each translation cluster can contain all of the plurality of potential translations that have a similar denotation and each of the plurality of translations that have a similar denotation can be included in a specific translation cluster. The technique can also include outputting, at the server, the one or more translation clusters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for aligning words in parallel translation sentences for use in machine translation.