Focusing on BLEU Can Bias Machine Translation Output
Though beam search can boost BLEU scores, it can also lead to high rates of misgendered pronouns, even when translating between two gendered languages. Continue reading
Advertise on Slator! Download the 2021 Online Media Kit Now
News and analysis of the latest developments in machine translation, computer-aided-translation, natural language processing, and other language-related areas in artificial intelligence.
Though beam search can boost BLEU scores, it can also lead to high rates of misgendered pronouns, even when translating between two gendered languages. Continue reading
Chinese tech company ByteDance (TikTok) hopes NeurST, which bypasses the transcription step integral to traditional, cascaded speech translation, will standardize benchmarks in the field. Continue reading
At SlatorCon Remote December 2020, Intento Co-founder and CEO Konstantin Savenkov discusses alternative pricing models that account for machine translation quality. Continue reading
Amazon launches Live Translation for Alexa from English into six languages; says new feature builds on previous tech, such as Amazon Translate. Continue reading
The most-cited neural machine translation research papers show how NMT came to dominate the field — and how academia and industry’s interests have evolved. Continue reading
November 2020 paper by University of Zurich and Lilt researchers measures how text presentation impacts translation quality via three tasks: copying, identifying errors, and revision. Continue reading
Google study shows human-paraphrased reference translations and new evaluation metric, BLEUP, produce better translations. Findings to be integrated into consumer-facing products. Continue reading
New research from DeepMind and Google explores end-to-end semi-automated dubbing; cites ongoing concerns around misuse of mimicry — from deep fakes to consent issues. Continue reading
MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) derives algorithm to handle lack of training data, making it easier to adopt current machine translation methods. Continue reading
The new machine translation model, trained on 2,200 language directions, builds on Facebook’s past research on parallel sentences not aligned with English. Continue reading
US and Chinese tech giants continue to churn out research on speech-to-text translation, machine translation for rare languages, and more. Time to check in. Continue reading
US Department of Homeland Security awards California startup in Phase 1 of RFP for “portable language translator.” Call for tenders still open until February 2021. Continue reading
A group of AI researchers takes the perennial human vs. machine debate to the next level by claiming their system outperforms professional translators...in news, on adequacy (read on for more caveats). Continue reading
Since its 2017 debut, Sockeye has powered Amazon Translate and gathered decent traction elsewhere. The new version adopts Gluon base code and gets a boost from Intel and NVIDIA. Continue reading
Facebook AI invites developers to improve a speech translation dataset and a open sources a toolkit that evaluates simultaneous speech and text translation. Continue reading
Google Translate is the most widely known and most used machine translation technology used globally. Google machine translation works by employing a neural machine translation (NMT) algorithm to do the hard work of translation instantly.
Different NMT solutions in the market use powerful language translation algorithms that are constantly being developed to suit the evolving needs of the translation and localization industry. These algorithms originated from the concept of deep learning. Read More
Deep learning seeks to artificially mimic the structure of the human brain’s neural networks in a computer. Neural networks control the brain’s thinking processes. By studying neural networks, scientists realized that once computers had become fast enough, and there was enough data on the internet, they could create an algorithm that trains large artificial neural networks. The wide adoption of neural machine translation has now resulted in the technology to be referred to as just machine translation.
Natural language machine translation is well on its way to becoming the single most important productivity enhancement technology for human translators. At the same time, fast improving quality means that stand-alone machine translation is rapidly finding new use cases.
In that sense, the pure play machine translation technology market relates to the development and deployment of machine translation technology without any human translation services. This differs from the human-in-the-loop models, such as post-editing of machine translation (where a human edits machine output) and interactive machine translation (where the machine provides prompts and the human accepts or modifies the suggested output).
The history of machine translation is divided into three main eras: Rules-based (1960s), Statistical (ca. 2007) and Neural (2016 onward). Academic research into neural machine translation (NMT) really only began in 2014, and NMT has since replaced SMT as the de facto standard in automated translation.
The first breakthrough in NMT was made through recurrent neural network (RNN) models, and eventually shifted into convolutional neural networks (CNN) with attention mechanisms. Today, the most popular model is the attention-based transformer, developed by Google.
Companies such as Google and Amazon Web Services have released free and open source neural libraries / frameworks for machine learning that further promote NMT research and use. Google, for instance, released the TensorFlow framework, a library for neural network-based machine learning applications. On top of that, they also released the widely used tensor2tensor neural library for sequence-to-sequence tasks, such as NMT.
Currently, approaches such as massive multilingual machine translation are being explored in order to remove the need for pivot languages (mostly English) and build more powerful models able to machine translate thousands of different language combinations.
Big Tech Machine Translation: Many of the big tech players have developed machine translation capabilities for their own purposes. Social platforms such as Facebook develop MT to make user-generated content accessible to other users in their own language and, more crucially, for their core business, to track content posted in hundreds of languages worldwide. Chinese e-commerce giant Alibaba, meanwhile, uses MT to solve the challenge of providing multilingual customer support to users of their products as well as making product information available in multiple languages. Typically, these solutions are not licensed to third-party corporates, although they theoretically could be.
Machine Translation Niche Players: Niche players are those companies whose main specialism is developing and deploying machine translation technology for licensing to enterprise customers, LSPs, and freelance linguists. They do not compete in the language services space or outside machine learning, and they are not research institutions. Examples of niche players in the machine translation space include the following:
Language Service Providers’ Machine Translation Offering: A fast growing number of LSPs offer human-in-the-loop services, where machine translation output is augmented by human translation / editing services and delivered to the customer. A number of LSPs develop proprietary MT solutions that are licensed to enterprise customers, third-party LSPs, and freelance linguists. These solutions are offered as stand-alone products or as part of a suite of products, without human translation services.
Machine Translation Research Community: The research community has always played an important role in the development of machine translation technology. Neural MT remained a fringe research topic until around 2016, when its adoption to the mainstream began in earnest. Since mid-2018, research activity on NMT published on the arXiv repository has soared. Key contributions to machine translation research come not only from academia, but also from national and independent research institutes and big tech, as well as a handful of LSPs / machine translation providers