Neural matching
A Google system that uses neural networks to understand how queries relate to page concepts, bridging vocabulary gaps between what users…
AlgorithmsBidirectional Encoder Representations from Transformers
BERT
A transformer language model Google applied to search from 2019, improving understanding of word order, prepositions and context within queries and passages.
In full
BERT reads a sequence bidirectionally, so it captures how 'to' and 'from' change a query's meaning — the classic example being a traveller query where 'to' determines the direction of travel. Google said it initially affected a substantial share of English queries and later expanded to all languages. It also improved featured snippet selection. As with other AI systems there is nothing to optimise for beyond writing naturally and precisely.
Example
'Can you get medicine for someone pharmacy' began returning results about collecting a prescription for another person rather than general pharmacy pages.
Related terms
RankBrain
Google's first machine-learning ranking system, introduced in 2015, which helps interpret how queries relate to concepts, especially…
MUM
Google's multimodal, multilingual model announced in 2021, able to transfer knowledge across languages and formats to answer complex…
Semantic search
Retrieval based on the meaning of a query and documents rather than literal keyword matching, using entities, context and learned…
Passage ranking
Google's ability to rank a page based on the relevance of a specific passage within it, even when the page overall is about something…
Where to read more
- Attention Is All You Need · Landmark Post · arxiv.org
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding · Landmark Post · arxiv.org
- Passage Re-ranking with BERT · Landmark Post · arxiv.org
- Understanding Searches Better Than Ever Before · Landmark Post · blog.google