PUBLICATIONS
The latest language news
Let us talk again about the latest language news. I have highlighted the news that I consider the most important.
Google introduces new AI model that translates vision, language into robotic operations
Robotics Transformer 2 (RT-2) is a groundbreaking vision-language-action model that brings us closer to a future of helpful robots. It has been trained on text and images from the web, allowing it to directly output robotic actions and effectively making it a capable “speaking robot.” Developing robots that can handle complex, abstract tasks in diverse and unfamiliar environments has been a challenging endeavor. Unlike chatbots, robots require real-world grounding and an understanding of their capabilities. Traditionally, this has meant the training of robots on billions of data points, which was time-consuming and impractical for most innovators. RT-2 takes a new approach to the problem. It improves robots’ reasoning abilities and eliminates the need for complex stacks of systems by enabling a single model to perform both complex reasoning and robot actions. Even with only a small amount of robot training data, RT-2 can transfer knowledge from its language and vision training data to direct a robot’s actions, even for tasks it has never been explicitly trained to do. The benefits of RT-2’s method are significant. It allows robots to rapidly adapt to novel situations and environments, performing as well as previous models on tasks in their training data and significantly outperforming them in unseen scenarios. Moreover, RT-2’s ability to transfer learned concepts to new situations brings robots closer to learning and adapting more like humans. This advancement not only signifies the convergence of AI and robotics but also holds immense promise for the development of more general-purpose robots that can better serve human-centered environments. While there is still much work to be done to fully realize the potential of helpful robots, RT-2 provides a glimpse of an exciting future for robotics—one where robots can learn from diverse data sources and tackle a wide array of tasks, bringing us closer to a world of advanced and capable robotic assistants.
Reading the mind with machines to help patients unable to speak
In Alexandre Dumas’s classic novel The Count of Monte-Cristo, a character named Monsieur Noirtier de Villefort suffers a terrible stroke that leaves him paralyzed. Though he remains awake and aware, he is no longer able to move or speak, relying on his granddaughter Valentine to recite the alphabet and flip through a dictionary to find the letters and words he requires. With this rudimentary form of communication, the determined old man manages to save Valentine from being poisoned by her stepmother and thwart his son’s attempts to marry her off against her will. Dumas’s portrayal of this catastrophic condition—in which, as he puts it, “the soul is trapped in a body that no longer obeys its commands”—is one of the earliest descriptions of locked-in syndrome. This form of profound paralysis occurs when the brain stem is damaged, usually because of a stroke but also as a result of tumors, traumatic brain injury, snakebite, substance abuse, infection, or neurodegenerative diseases like amyotrophic lateral sclerosis (ALS). The condition is thought to be rare, though just how rare is hard to say. Many locked-in patients can communicate through purposeful eye movements and blinking, but others can become completely immobile, losing their ability even to move their eyeballs or eyelids, rendering the command “blink twice if you understand me” moot. As a result, patients can spend an average of 79 days imprisoned in a motionless body, conscious but unable to communicate, before they are properly diagnosed. The advent of brain-machine interfaces has fostered hopes of restoring communication to people in this locked-in state, enabling them to reconnect with the outside world. These technologies typically use an implanted device to record the brain waves associated with speech and then use computer algorithms to translate the intended messages. The most exciting advances require no blinking, eye tracking, or attempted vocalizations but instead capture and convey the letters or words a person says silently in their head. Using this technology, researchers have recorded hours of data and fed it into sophisticated machine learning algorithms. Today, almost 40 people worldwide have been implanted with microelectrode arrays, with more coming on line. Another approach by Jun Wang of the University of Texas at Austin uses an advanced imaging technique called magnetoencephalography (MEG), which records magnetic fields on the outside of the skull that are generated by the electric currents in the brain, and then translates the signals into text. Right now, he is trying to build a device that uses MEG to recognize the 44 phonemes, or speech sounds, in the English language—like ph or oo—which could be used to construct syllables, then words, then sentences.
ModernMT introduces Trust Attention to boost MT quality
https://blog.modernmt.com/modernmt-introduces-trust-attention-to-improve-mt-quality/
ModernMT has introduced Trust Attention, which allows MT engines to prioritize more trustworthy data and have this data influence model behavior more heavily. The new innovation provides global enterprises with a superior platform to build highly tuned enterprise-specific translation engines. This progress is especially noted and clear with dynamically adaptive MT models like ModernMT, where small amounts of ongoing corrective expert feedback result in the continuous improvement of the quality of MT output. ModernMT’s historical track record has been so impressive that it does not seem unreasonable to point out that ModernMT’s performance across billions of samples and many languages is approaching the singularity in production-use scenarios. This is the point at which human editors are unable to tell whether the sample is coming from a human or machine since they are so close in quality and style. MT technology continues to evolve and improve, with recent updates providing much richer and more granular document-level contextual awareness. In the summer of 2023, we are at an interesting junction in the development of AI-based language translation technology, where we now see that Large Language Models are also an emerging technological approach to having machines perform the task of language translation. LLMs are particularly impressive in handling idioms and enhancing the fluency of machine translations. The AI product team at Translated continues to research and investigate the possibilities for the continued improvement of pure MT models, hybrid MT and general AI models, as well as pure general AI models. ModernMT Version 7 introduces a significant upgrade to its core adaptive MT system. This new version introduces Trust Attention, a novel technique inspired by how human researchers prioritize information from trusted sources, and the Version 7 model preferentially uses identified trustworthy data both in training and inference. While it is common practice in the industry to use automated algorithm-driven methods to drive data validation and verification practices, Translated’s 20 years of experience working with human translators show that human-verified data is the most trustworthy data available to drive the learning of language AI models. This foundation of human-verified data is the most influential driver of preferential learning in the models of ModernMT Version 7. “Garbage in, garbage out” is a concept in computing and AI that highlights the importance of the quality of input data. It means that if the data input to a system such as an AI model or algorithm is of poor quality, is inaccurate, or is irrelevant, the system’s output will also be of poor quality, be inaccurate, or be irrelevant. This concept is particularly significant in the context of AI models which use machine learning and deep learning models and which rely heavily on the data used for training and validation. If the training data is biased, incomplete, or contains errors, the AI model will likely produce unreliable or biased results. Traditional MT systems are generally not able to distinguish between trustworthy data and lower-quality training material during the training process, and typically all the data has equal weight. Thus, high-quality data and high-volume noisy data can have essentially the same amount of impact on how a translation model will perform. Trust Attention allows an engine to prioritize more trustworthy data and have this data influence model behavior more heavily.
Machine learning mitigates the spread of fake news
https://www.kompas.id/baca/english/2023/08/02/en-mesin-pembelajar-memitigasi-penyebaran-kabar-bohong
The development of digital technology has made it easy for false information to circulate. Machine learning or learning machines with block chain technology or blockchain can be used to help mitigate the spread of fake news. The latest research at the State University of New York at Binghamton, United States, is developing a study offering tools to recognize the patterns of false information. This helps content creators detect inaccuracies that occur. The research proposes a machine learning system—a part of artificial intelligence—which uses data and algorithms to mimic the way humans determine if content may harm readers. An example of this is the information touting fake alternative treatments during the height of the COVID-19 pandemic. The machine learning framework uses data and algorithms to identify indicators of misinformation. It then uses these examples to improve the detection process. Based on the information gathered, machine learning systems can help mitigate fake news by distinguishing which messages are likely to be the most damaging if allowed to spread. The researchers propose a survey of 1,000 people from two groups, namely fake news checkers and content users who may be exposed to fake news messages. The survey will describe the existing blockchain system and measure participants’ willingness to use the system in different scenarios. The spread of false news must be prevented because it has the potential to lead to negative impacts. The machine learning system will be used to analyze text and generate scores representing the likelihood of each article containing fake news.
Machine Translation Specialist, R&D
Artem Eidel
Artem Eidel is a Machine Translation Specialist in R&D. His responsibilities include creating and maintaining MT models, evaluating MT quality, and selecting suitable MT providers for different language pairs and subjects.
