Fast Attention-based Learning-To-Rank Model for Structured Map Search

Recent works show that Transformer-based learning-to-rank (LTR) approaches can outperform previous well-established ranking methods, such as gradient-boosted decision trees (GBDT), on document and passage re-ranking problems. A common assumption in these works is that the query and the result documents are comprised of purely textual information without explicit structure. In map search, the relevance of results is determined based on rich heterogeneous features - textual features derived from the query and the results, geospatial features such as proximity of a result to the user, structured features reflecting the address format of the result, and the perceived structure of the query. In this work, we propose a novel deep neural network LTR architecture, capable of seamlessly handling heterogeneous inputs, similar to GBDT-based methods. At the same time, unlike GBDT, the architecture does not require human input via (numerous) carefully-crafted features. Instead, features are inferred through a self-attention mechanism. Our model implements two lightweight attention layers optimized for ranking: the first layer computes query-result similarities, the second implements listwise ranking inference. We perform evaluation on several single language and one multilingual dataset. Our model outperforms by a wide margin other Transformer-based ranking architectures and has equal or better performance than GBDT models. Equally important, runtime inference is orders of magnitude faster than other Transformer architectures, significantly reducing hardware serving costs. The model is a low-cost alternative suitable to power ranking in industrial map search engines across a variety of languages and markets.