A video from How does Google MUM work? The previous natural language processing model called BERT could also understand things and entities, but could only provide short answers or facts in the sense of numbers, names or data. MUM can do much more than that , it is based on the T ext T ext T ransfer T ransformer , also called T5. Pretty jokers on Google, right? The C4 corpus is two instances larger than Wikipedia, i.e. Wikipedia squared. That’s quite a lot of text for MUM to learn from.
To put it simply While
T5 can translate (green), evaluate linguistic Egypt WhatsApp Number List acceptability (red), evaluate the similarity of sentences in numbers (yellow) and also write complete summaries (blue). BERT could only answer individual chunks of words, such as “300 meters high”, “Thomas Müller” or other short answers, MUM can now answer entire sentences, evaluate them, translate them correctly and also summarize complete texts in one sentence. Multilingual application possible The transfer into other languages in particular seems to be of great benefit to Google.
Search in videos possible
So every question can be translated into A video from Japanese, Google Brazil WhatsApp Number List gets the answer there, summarizes the text and translates it back into German. Google MUM also works across text and can analyze videos very well. These search queries look groundbreaking to me. Another example shows a sweater with a striking pattern. The user holds the camera on it and writes “socks with this pattern,” whereupon Google presents them with socks with this pattern.