Event
BERT untersucht Sprachkontext
In 2018 a research team introduced BERT, a pretrained model based on the Transformer architecture for language-understanding tasks. Its word representations use context from both directions. One pretraining task predicted masked text from its surrounding context; the model could then be fine-tuned for specific tasks. Google announced the release of code and pretrained models in November. Source
According to Google, the models initially released in November 2018 were English-only. In October 2019 the company described a separate application of BERT models in its search engine: for ranking, it expected improved understanding of about one in ten English-language searches in the US. This was a company claim at the time, not a worldwide success rate. Research release and Search announcement
Sources
Forschungsfreigabe und Suchanwendung: Google-Mitteilungen 2018 und 2019 mit begrenzter Anbieterzahl.
Forschungsarbeit und Google-Ankündigung: Modellansatz und Freigabe.
Offen
Unabhängige Untersuchungen zur tatsächlichen Wirkung auf Suchergebnisse und zu weiteren Sprachen prüfen. / Investigate independent evidence about actual effects on search results and additional languages.