Facebook, Inc.

International Business Machines Corporation

Google Inc.

Fei Huang

Ying Zhang

Kay Rottmann

Necip Fazil Ayan

Matthias Gerhard Eck

Xiaolong Wang

Jian-ming Xu

Diego Melendo Casado

Jakob Nicolaus Foerster

Francoise Beaufays

Fuchun Peng

Lars Seren Backstrom

Thomas Frederick Dimson

Justin Moore

Aditya Pal

Stanislav Funiak

Jan Kalis

Daniel Matthew Merl

Taylor Gordon

Feng Liang

Jian-Ming Xu

Seyoung Park

A translation text output corresponding to a machine translation from a source text of a first language to a target text of a second language is received. The translation text output includes at least one sentence including at least one translation chunk. A first translation quality confidence estimate for the at least one sentence of the text and a second translation quality confidence estimate for the at least one translation chunk are received. A first translation confidence indicator for the at least one sentence and a second translation confidence indicator for the at least one translation chunk are displayed. The first translation confidence indicator is determined from among a plurality of different first translation confidence indicators based upon the first translation quality confidence estimate. The second translation confidence indicator is determined from among a plurality of different second translation confidence indicators based upon the second translation quality confidence estimate.

System and method for adaptive quality estimation for machine translation post-editing

A computer-implemented method for leveraging computer aided input segmentations is provided. The computer aided input segmentations identify divisions in a source text. The division are utilized to translate the source text from a first language to a second language. In this regard, segmentation boundaries are designated within the source text based on the computer aided input segmentations, which determine the divisions. Each division is automatically translated into the second language to generate translated segments corresponding to the division. These translated segments are combined to generate a translated text.

Computer aided input segmentation for machine translation

Exemplary embodiments relate to techniques for improving a machine translation system. The machine translation system may include one or more models for generating a translation. The system may generate multiple candidate translations, and may present the candidate translations to different groups of users, such as users of a social network. User engagement with the different candidate translations may be measured, and the system may determine which of the candidate translations was most favored by the users. For example, in the context of a social network, the number of times that the translation is liked or shared, or the number of comments associated with the translation, may be used to determine user engagement with the translation. The models of the machine translation system may be modified to favor the most-favored candidate translation. The translation system may repeat this process to continue to tune the models in a feedback loop.

Optimizing machine translations for user engagement

Systems, methods, and non-transitory computer-readable media can generate a set of candidate content items from a plurality of content items that are available in the social networking system for a first user. A corresponding score for each of the candidate content items can be generated based at least in part on one or more social affinity coefficients corresponding to the first user and a respective second user associated with a candidate content item, wherein a social affinity coefficient provides a quantitative measurement of the strength of a relationship between two users. A first set of content items from the set of candidate content items can be determined based at least in part on the respective scores, wherein content items in the first set are included in a content feed provided to the first user.

Systems and methods for providing content

Snippets can be represented in a language-independent semantic manner. Each portion of a snippet can be represented by a combination of a semantic representation and a syntactic representation, each in its own dimensional space. A snippet can be divided into portions by constructing a dependency structure based on relationships between words and phrases. Leaf nodes of the dependency structure can be assigned: A) a semantic representation according to pre-defined word mappings and B) a syntactic representation according to the grammatical use of the word. A trained semantic model can assign to each non-leaf node of the dependency structure a semantic representation based on a combination of the semantic and syntactic representations of the corresponding lower-level nodes. A trained syntactic model can assign to each non-leaf node a syntactic representation based on a combination of the syntactic representations of the corresponding lower-level nodes and the semantic representation of that node.

Language independent representations

A machine translation system can improve results of machine translations by employing preferred translations, such as human translated phrases. In some implementations, the machine translation system can use the preferred translations as heavily weighted training data when building a machine translation engine. In some implementations, the machine translation system can use the preferred translations as an alternate to a result that would have otherwise been produced by a machine translation engine. While it is infeasible to obtain human translations for all translation phrases, preferred translations can be used for problem phrases for which machine translation engines often produce poor translations. The machine translation system can identify problem phrases by assigning a quality score to each translation in a set of translations. The machine translation system can identify, as the problem phrases, n-grams that appear with a frequency above a frequency threshold in translations with quality scores below a threshold.

User feedback for low-confidence translations

The described technology can provide semantic translations of a selected language snippet. This can be accomplished by mapping snippets for output languages into a vector space; creating predicates that can map new snippets into that vector space; and, when a new snippet is received, generating and matching a vector representing that new snippet to the closest vector for a snippet of a desired output language, which is used as the translation of the new snippet. The procedure for mapping new snippets into the vector space can include creating a dependency structure for the new snippet and computing a vector for each dependency structure node. The vector computed for the root node of the dependency structure is the vector representing the new snippet. A similar process is used to train a transformation function for each possible node type, using language snippets already associated with a dependency structure and corresponding vectors.

Deep translations

Technology is disclosed for snippet pre-translation and dynamic selection of translation systems. Pre-translation uses snippet attributes such as characteristics of a snippet author, snippet topics, snippet context, expected snippet viewers, etc., to predict how many translation requests for the snippet are likely to be received. An appropriate translator can be dynamically selected to produce a translation of a snippet either as a result of the snippet being selected for pre-translation or from another trigger, such as a user requesting a translation of the snippet. Different translators can generate high quality translations after a period of time or other translators can generate lower quality translations earlier. Dynamic selection of translators involves dynamically selecting machine or human translation, e.g., based on a quality of translation that is desired. Translations can be improved over time by employing better machine or human translators, such as when a snippet is identified as being more popular.

Predicting future translations

Technology is disclosed for creating and tuning classifiers for language dialects and for generating dialect-specific language modules. A computing device can receive an initial training data set as a current training data set. The selection process for the initial training data set can be achieved by receiving one or more initial content items, establishing dialect parameters of each of the initial content items, and sorting each of the initial content items into one or more dialect groups based on the established dialect parameters. The computing device can generate, based on the initial training data set, a dialect classifier configured to detect language dialects of content items to be classified. The computing device can augment the current training data set with additional training data by applying the dialect classifier to candidate content items. The computing device can then update the dialect classifier based on the augmented current training data set.

Machine learning dialect identification

A likely source language of a media item can be identified by attempting an initial language identification of the media item based on intrinsic or extrinsic factors, such as words in the media item and languages known by the media item author. This initial identification can generate a list of most likely source languages with corresponding likelihood factors. Translations can then be performed presuming each of the most likely source languages. The translations can be performed for multiple output languages. Each resulting translation can receive a corresponding score based on a number of factors. The scores can be combined where they have a common source language. These combined scores can be used to weight the previously identified likelihood factors for the source languages of the media item.

Universal translation

A system for identifying language(s) for content items is disclosed. The system can identify different languages for content item words segments by identifying segment languages that maximize a probability across the segments. The probability can be a combination of: an author's likelihood for the language identified for the first word; a combination of transition frequencies for selected languages identified for words, the transition frequencies indicating likelihoods that a transition occurred to the selected language from the previous word's language; and a combination of observation probabilities indicating, for a given word in the content item, a likelihood the given word is in the identified language. For an in-vocabulary word, the observation probabilities can be based on learned probability for that word. For an out-of-vocabulary word, the probability can be computed by breaking the word into overlapping n-grams and computing combined learned probabilities that each n-gram is in the given language.

Identifying multiple languages in a content item

A confidence scoring system can include a model trained using features extracted from translations that have received user translation ratings. The features can include, e.g. sentence length, an amount of out-of-vocabulary or rare words, language model probability scores of the source or translation, or a semantic similarity between the source and a translation. Parameters of the confidence model can then be adjusted based on a comparison of the confidence model output and user translation ratings, where the user translation ratings can be selected or weighted based on a determination of individual user fluentness. After the confidence model has been trained, it can produce confidence scores for new translations. If a confidence score is higher than a threshold, it can indicate the translation should be selected for automatic presentation to users. If the confidence score is below another threshold, it can indicate the translation should be updated.

Growing community of inventors

Boonton, NJ, United States of America

Fei Huang