The machine learning service text translation capability has been upgraded, and the Chinese literal translation model makes the translation expression more authentic!

The HMS Core machine learning service text translation capability provides translation services in multiple languages and application scenarios. For example, in the scenario of traveling abroad, users can use the voice translation broadcast function of the app to communicate in scenarios such as taxis and hotel check-ins. , you can also read restaurant menus, street sign information, etc. through the photo translation function.

Chinese literal translation model upgrades text translation capabilities

Most of the current mainstream translation models use English, which is rich in corpus resources, as an intermediate language for "bridging" translation. However, after the English transfer, the translation accuracy is lost, and the effect of doubling the computing resources is reduced. In order to respond to the "Belt and Road" initiative and facilitate the process of globalization with diversity and openness, the language translation requirements for domestic overseas applications are clear and require system localization deployment. The translation quality requirements for some language directions such as Chinese, Japanese, Chinese and Russian are relatively high. The translation effect of language direction continued to improve.

Based on this, the HMS Core machine learning service has upgraded the text translation capabilities , and the Chinese literal translation models for Chinese, Japanese, German, French, and Russian languages have been launched in the new version. Compared with English bridge translation, the Chinese literal translation model can achieve 300 characters per second, the end-to-end translation speed is less than 150ms, the translation delay is reduced by 100%, and the translation speed is faster; the translation results for Chinese characteristic words and slang are more authentic, Better translation quality. The upgraded text translation capabilities can meet the needs of Chinese companies going overseas or foreign companies entering China and other scenarios with higher translation requirements.

The Chinese literal translation model scheme also participated in the sub-task of the WMT2021 International Machine Translation Competition (Shared Task: Triangular MT: Using English to improve Russian-to-Chinese machine translation), and won the first place in this task with a significant advantage.

Comparison of Chinese literal translation and bridging translation

French translation

【Original】Smart Launcher est un lanceur pour d'applications Android qui substitue l'interface classique de votre téléphone avec une autre qui est un peu plus simple et qui vous permettra d'accéder à toutes vos applications plus rapidement et plus confortablement.

【English Bridge】Smart Launcher is a launcher for Android applications, which can replace the classic interface of your phone with a simpler interface, allowing you to access all applications faster and more comfortably.

【英文原文】Smart Launcher is an Android app launcher that replaces the classic interface of your phone with another simpler interface that will give you faster and more comfortable access to all your apps.

Russian to Chinese

[Original] о, ладно о, когда кто кто грубит тебе в коридоре является, без сомнений, самой меньшей из тех проблем проблем проблем проблем которыми которыми сталкиваешься в старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших старших которыми 【【меньшей

[English bridging] Oh please, being rude to you in the hallway is without a doubt the least important issue you faced in high school.

【英文翻译】Oh please, when someone is rude to you in the hallway is definitely the smallest problem you ever had in high school.

German to Chinese

【Original】Dieser von klassischen Laufschuhen inspirierte Herren-Sneaker zeichnet sich durch einen Materialmix aus Wildleder und verschiedenen Gewebearten aus und präsentiert sich in der Cruise 2020 Kollektion mit einer neuen Mid-Top-Silhouette, die genau bis zum Knöchel reicht.

【English Bridge】Inspired by classic running shoes, these men's sneakers feature a mix of suede and different fabrics, and feature a new mid-top silhouette that stretches all the way to the ankle in the Cruise 2020 collection.

【英文翻译】Inspired by classic running shoes, this men's sneaker features a new mid-cut silhouette in the Cruise 2020 collection with a mix of suede and different fabrics, extending fully to the ankle.

Technical advantages of Chinese literal translation model

The Chinese literal translation model draws on Huawei's latest research results in machine translation. By using Russian-English and English-Chinese corpus for knowledge distillation, in the case of a small amount or no Russian-Chinese corpus, combined with explicit course learning and training strategies, it can get good quality. The Russian-Chinese translation model solves the problems of low resources and cold start of the English bridging translation system.

Chinese literal translation system

Technical point 1: Multi-lingual Encoder Decoder Enhancement

Taking the Russian-Chinese translation task as an example, through the multilingual model, the English-Chinese corpus is introduced, combined with knowledge distillation, to strengthen the decoding ability of the decoder layer on the translation side; the introduction of Russian-English materials to strengthen the encoding ability of the encoder layer on the original side. So as to solve the cold start problem to a great extent.

Technical point 2: Explicit Curriculum Learning for Denoising

During the training process, combined with explicit curriculum learning skills, a three-stage incremental training program is divided according to the amount of noise. In the first stage, the full amount of noisy corpus is used for training, so that the model converges quickly; in the second stage, incremental training is performed after denoising with a bilingual alignment tool; in the third stage, the enhanced denoised bilingual data is used for incremental training. Complete the final convergence of the model.

Technical Point 3: Data Augmentation Scheme Based on Combination Decoding Strategy (FTST For Data Augmentation)

The full name of FTST is Forward Translation and Sampling Backward Translation. This technology uses sampling in the Backward Model for data enhancement; in the Forward Model, beam search is used for data balance. In comparative experiments, the best results were obtained.

In addition to the existing four languages, the machine learning service text translation capability plan will be launched this year in Japan, France, Germany, Russia, Korea, Portuguese, Spanish, Turkish, Thai, Arabic, Malay, Italian, Polish, Dutch, Vietnamese There are Chinese literal translation models for 15 countries in total, basically covering major countries and regions in the world such as China, Asia, Africa and Latin America, Europe, and South America.

Learn more details>>

Visit the official website of Huawei Developer Alliance
Get development guidance documents
Huawei Mobile Services Open Source Warehouse Address: GitHub , Gitee

Follow us to know the latest technical information of HMS Core for the first time~