Facebook, Inc.

Meta Platforms, Inc.

Matthias Gerhard Eck

Ying Zhang

Kay Rottmann

Juan Miguel Pino

Fei Huang

Necip Fazil Ayan

Kay Rottman

Rui Andre Augusto Ferreira

Priya Goyal

Lukasz Langa

Alexander Waibel

William Arthur Hughes

Yury Andreyevich Zemlyanskiy

Juan Sebastian Rassa Robayo

According to examples, a system for automatically optimizing thresholds of content processing models that select content for presentation to users may include a processor and a memory storing instructions. The processor, when executing the instructions, may cause the system to select a subset of the content processing models for a content policy grouping. The subset of content processing models comprises models selected from a plurality of content processing models based on content rejection rates and models that are selected based on corresponding model probabilities. The system may further obtain an optimized threshold for each model of the subset of content processing models based on an iterative global optimization technique. The system may thereby facilitate automatic selection or rejection of the content pieces for presentation to users on an online system based on the policies associated with corresponding content policy grouping by employing the subset of content processing models with the optimized thresholds.

Model thresholds for digital content management and selection

The present disclosure is directed toward systems and methods for providing translations of electronic messages via a social networking system. For example, systems and methods described herein involve determining whether to provide an electronic message or a translation of the electronic message to a recipient based on social networking activities of the recipient. Furthermore, systems and methods described herein can provide a translation of an electronic message based on an analysis of social networking activities of one or more recipients of the electronic message.

Providing translations of electronic messages via a social networking system

Technology is disclosed for building correction models that correct natural language snippets. Correction models can include rules comprising pairs of word sequences identified from viable correction snippet pairs, where a first sequence of words in the pair should be replaced with a second sequence of words in the pair. Viable correction snippet pairs can be identified from among pairs of language snippets, such as a post to a social media website and a subsequent update to that post. Viable corrections can be the snippet pairs that both have no more unaligned words than a word alignment threshold and have no aligned word pair with a character edit difference above an edit distance threshold. In some implementations, word alignments can be found by aligning all the characters between a pair of language snippets, and identifying aligned words as those that have at least one aligned letter in common.

Machine-translation based corrections

Exemplary embodiments relate to techniques for improving machine translation systems. The machine translation system may apply one or more models for translating material from a source language into a destination language. The models are initially trained using training data. According to exemplary embodiments, supplemental training data is used to train the models, where the supplemental training data uses in-domain material to improve the quality of output translations. In-domain data may include data that relates to the same or similar topics as those expected to be encountered in a translation of material from the source language into the destination language. In-domain data may include material previously translated from the source language into the destination language, material similar to previous translations, and destination language material that has previously been the subject of a request for translation into the source language.

Language model using reverse translations

Exemplary embodiments relate to detecting, removing, and/or replacing objectionable words and phrases in a machine-generated translation. A classifier identifies translations containing target words or phrases. The classifier may be applied to the output translation to remove target words and phrases from the translation, or to prevent target words and phrases from being automatically presented. Further, the classifier may be applied to a translation model to prevent the target words and phrases from appearing in the output translation. Still further, the classifier may be applied to training data so that the translation model is not trained using the target words of phrases. The classifier may remove target words or phrases only when the target words or phrases appear in the output translation but not the source language input data. The classifier may be provided as a standalone service, or may be employed in the context of a machine translation system.

Target phrase classifier

Exemplary embodiments provide techniques for evaluating when words or phrases of a translation were generated with a low degree of confidence, and conveying this information when the translation is presented. For example, if a source language word is encountered in source material for translation, but the source language word was only encountered a few times (or not at all) in the training data used to train the translation system, then the resulting translation may be flagged as being of low confidence. Other situations, such as the generation of two equally-likely translations, or translation system model disagreement, may also indicate a questionable translation. When the translation is displayed, questionable words and phrases may be flagged, and possible alternative translations may be presented. If one of the alternatives is selected, this information may be used to update the translation system's models in order to improve translation quality in the future.

Identifying risky translations

Machine translation system employing classifier

Specialized language processing engines can use author-specific or reader-specific language models to improve language processing results by selecting phrases most likely to be used by an author or by tailoring output to language with which the reader is familiar. Language models that are author-specific can be generated by identifying characteristics of an author or author type such as age, gender, and location. An author-specific language model can be built using, as training data, language items written by users with the identified characteristics. Language models that are reader-specific can be generated using, as training data, language items written by or viewed by that reader. When implementing a specialized machine translation engine, multiple possible translations can be generated. An author-specific language model or a reader-specific language model can provide scores for possible translations, which can be used to select the best translation.

Language model personalization

Technology is disclosed that improves language coverage by selecting sentences to be used as training data for a language processing engine. The technology accomplishes the selection of a number of sentences by obtaining a group of sentences, computing a score for each sentence, sorting the sentences based on their scores, and selecting a number of sentences with the highest scores. The scores can be computed by dividing a sum of frequency values of unseen words (or n-grams) in the sentence by a length of the sentence. The frequency values can be based on posts in one or more particular domains, such as the public domain, the private domain, or other specialized domains.

Data sorting for language processing such as POS tagging

Technology is disclosed for mining training data to create machine translation engines. Training data can be mined as translation pairs from single content items that contain multiple languages; multiple content items in different languages that are related to the same or similar target; or multiple content items that are generated by the same author in different languages. Locating content items can include identifying potential sources of translation pairs that fall into these categories and applying filtering techniques to quickly gather those that are good candidates for being actual translation pairs. When actual translation pairs are located, they can be used to retrain a machine translation engine as in-domain for social media content items.

Mining multi-lingual data

Technology is discussed herein for identifying trending actions within a group of posts matching a query. A group of posts can be selected based on specified actions, action targets, or parameters such as author age, location, gender, when the posts were posted or what keywords they contain. Selected posts can be divided into sentences and a dependency structure can be created for each sentence classifying portions of the sentence as actions or action targets. Statistics can be generated for each sentence or post indicating whether it matches the actions, action targets, or other parameters specified in the query. Based on these statistics, additional information can be gathered to respond to questions posed in the query.

Analyzing language dependency structures

Technology is discussed herein for identifying comparatively trending topics between groups of posts. Groups of posts can be selected based on parameters such as author age, location, gender, etc., or based on information about content items such as when they were posted or what keywords they contain. Topics, as one or more groups of words, can each be given a rank score for each group based on the topic's frequency within each group. A difference score for selected topics can be computed based on a difference between the rank score for the selected topic in each of the groups. When the difference score for a selected topic is above a specified threshold, that selected topic can be identified as a comparatively trending topic.

Growing community of inventors

San Francisco, CA, United States of America

Matthias Gerhard Eck