Google has developed a new search algorithm, SMITH, that can understand long queries and documents better than BERT. The information about it was published by Google in a research paper.
The new SMITH model is designed to semantically match long texts. While BERT is trained to understand words in the context of sentences, SMITH tries to understand whole documents. That is, the model focuses on understanding fragments in the context of the whole document.
Unlike BERT which is trained on datasets to predict randomly hidden words from context in sentences - SMITH is trained to predict the next block of sentences.
This training will help the new search algorithm better understand lengthy documents.
Whether SMITH will be fully utilised in Google's work is not yet known. The company has not made any official statements on this matter.