Researchers are using artificial intelligence in new, creative ways. Some use cases – predicting earthquakes, recognizing objects, and comparing medical scans, for example – offer tantalizing glimpses of a promising future. Language translation has long been considered a natural application for AI, but processes are ripe for improvement. Now, MIT researchers have announced an “unsupervised” language translation model whose early returns signal a future with “faster, more efficient computer-based translations of far more languages.”
Most widely-used translation programs (like Google’s) learn via models. These programs are trained to “look for patterns in millions of documents – such as legal and political documents, or news articles” that have been translated from one language to another by humans. Once they identify a new word in one language, they can find its match in a different language through pattern analysis.
This approach, however, is labor intensive and inefficient – partially due to its reliance on specific translations from one language to another to accurately translate words. New “monolingual” models have attempted to rectify these issues by translating “without direct translational information between the two,” but are slow and require significant computing power to work.
To combat these pain points, MIT researchers with the Computer Science and Artificial Intelligence Laboratory (CSAIL) developed a new technique. By employing a statistical metric commonly used for pixel alignment called the Gromov-Wasserstein distance – which “essentially measures distances between points in one computational space and matches them to similarly distanced points in another space” – in conjunction with a vectorized system called “word embeddings”, where “words of similar meanings [are] clustered closer together,” researchers were able to create a system that can deduce likely direct translations via the relative distances of words within each vector.
Using relational distances negates the time-consuming, laborious process of creating perfect word alignments. Gromov-Wasserstein is “tailor-made” for this purpose, said CSAIL Ph.D. student David Alvarez-Melis, who was the first author of the paper presenting the findings. “If there are points, or words, that are close together in one space, Gromov-Wasserstein is automatically going to try to find the corresponding cluster of points in the other space,” explained Alvarez-Melis.
Alvarez-Melis used months of the year as an example: despite differences between languages, the MIT-developed model “would see a cluster of 12 vectors that are clustered in one embedding and a remarkably similar cluster in the other embedding.” That alignment – “The model doesn’t know these are months,” clarified Alvarez-Melis – allows for simultaneous alignment of an entire vector space.
The result is accuracy on par with or exceeding existing monolingual models, but with greater speed and significantly less operating power. It’s a tremendously exciting step towards the goal of truly unsupervised word alignment, and another example of the power of AI being fine-tuned and deployed in new, creative, and useful ways.
If You’re Wondering When A.I. Will Start Making Market Predictions…
Guess what – it already is. Hedge funds and large institutional investors have been using Artificial Intelligence to analyze large data sets for investment opportunities, and they have also unleashed A.I. on charts to discover patterns and trends. Not only can the A.I. scan thousands of individual securities and cryptocurrencies for patterns and trends, and it generates trade ideas based on what it finds. Hedge funds have had a leg-up on the retail investor for some time now.
Not anymore. Tickeron has launched a new investment platform, and it is designed to give retail investors access to sophisticated AI for a multitude of functions:
And much more. No longer is AI just confined to the biggest hedge funds in the world. It can now be accessed by everyday investors. Learn how on Tickeron.com.