Google DeepMind researchers are exploring whether one AI model could find and rank search results, replacing a process often split between two systems.

The approach, called Autoregressive Ranking, generates identifiers for documents to produce a ranking. The team developed a training method called SToICaL to teach it which results should appear higher. The paper explains the approach.

A similar two-stage workflow is already used in Google Cloud’s developer tools. Applications first retrieve likely matches, then send those documents to a separate ranking API. Google says this second step assesses how well each document answers the query, beyond the similarity checks used for initial matching. The API can also select content for an AI model to use when generating an answer.

Tests on WordNet, a database of word relationships, showed performance comparable to a more computationally expensive ranking model.

The shopping results were mixed. One version improved the overall ranking score from 95.23 to 97.21 out of 100, while accuracy at selecting the top result fell from 95.16% to 70.32%.

That comparison covered 310 evaluation examples. The target rankings were generated using another model’s similarity scores, so the results measure agreement with that benchmark.

The experiments supplied the model with candidate documents in its prompt; they did not demonstrate a replacement for Google’s full web-search infrastructure.

The paper, coauthored with researchers at UMass Amherst and UT Austin, was first submitted in January and last revised in February. It does not announce a rollout in Google Search.

The post Google researchers improved AI search and made the top result worse appeared first on Search Engine Watch.

Source:: searchenginewatch.com