At Microsoft, we are always looking for ways to empower our customers to achieve more by delivering our most advanced AI-enabled services. In March 2021, we launched the preview release of semantic search on Azure Cognitive Search, which allows our customers’ search engines to retrieve and rank search results based on the semantic meaning of search keywords rather than just their syntactical interpretation. We introduced this functionality by leveraging state-of-the-art language models that power Microsoft Bing search scenarios across several languages – a result of the recent advancements in developing large, pretrained transformer-based models as part of our Microsoft AI at Scale initiative.
Today, we are excited to announce that we are extending these capabilities to enable semantic search across multiple languages on Azure Cognitive Search.
Semantic search consists of three scenarios – semantic ranking, captions and answers – and customers can easily enable them via the REST API or Azure Portal to get semantic search results. The following examples illustrate how these scenarios are being delivered across different languages, where we rank search results based on our semantic ranker, followed by extracting and semantically highlighting the answer to the search query.
The language models powering semantic search are based on our state-of-the-art Turing multi-language model (T-ULRv2) that enables search across 100+ languages in a zero-shot fashion. Using global data from Bing, these models have been fine-tuned across various tasks to enable high-quality semantic search features for multiple languages and have been distilled further to optimize for serving real-world online scenarios at a significantly lower cost. Below is a list of the various innovations that are powering semantic search today.
Since their introduction, the models have been serving Bing search traffic across several markets and languages, delivering high-quality semantic search results to Bing users worldwide. Additionally, we have validated the quality of semantic ranking on Azure Cognitive Search using a variety of cross-lingual datasets – these include academic benchmark datasets (e.g. XGLUE web page ranking) as well as real-world datasets from services currently powered by Azure Cognitive Search (e.g. Microsoft Docs). Our results showed several points of gain in search relevance metrics (NDCG) over the existing BM25 ranker for various languages such as French, German, Spanish, Italian, Portuguese, Chinese and Japanese. For semantic answers, our evaluations were based on multiple datasets focused on Q&A tasks. Current academic benchmark leaderboards for Q&A scenarios measure accuracy of answer extraction for a given passage. However, our assessments were required to go a step further and consider more real-world intricacies involving multiple steps (see Figure 3) to extract an answer from a search index: (1) documents retrieval from the search index, (2) candidate passage extraction from the given documents, (3) passage ranking across candidate passages, and (4) answer extraction from the most relevant passage. We observed that our model accuracy for French, Italian, Spanish and German languages is equivalent to that of English language.
The following table summarizes the set of languages and queryLanguage parameter values that we currently support via the REST API to enable semantic search on Azure Cognitive Search. Note that we have also added speller support for Spanish, French and German languages. For languages marked as “preview”, we encourage you out to try the capability for your search index and give us your feedback. For detailed instructions on how to configure semantic search for your target language, please refer to our documentation.
With additional support for new languages, we are very excited to extend access to our state-of-the-art AI-enabled search capabilities to developers and customers worldwide. Please sign up for our preview to try out semantic search today!
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.