Files Search Relevance using Wide and Deep Networks
Abstract
Unlike web-search, enterprise search queries are more contextual, documents are private and heterogeneous, and ground-truth relevance labels are scarce. Production systems at this scale have long relied on gradient-boosted decision tree (GBDT) models such as LightGBM with hand-engineered features, given their speed, interpretability, and robustness on numeric features. However, high-quality semantic embeddings from large, pre-trained language models open new possibilities for neural learning-to-rank models that can jointly learn from sparse hand-crafted features (e.g., personalization signals such as user activity, document recency, and lexical matching features like BM25) and dense semantic representations. In this paper, we present our work on using wide & deep neural networks for search relevance over the enterprise files data. Due to privacy constraints on customer data, which limits direct inspection, and the lack of publicly available enterprise search datasets, we complement our private data experiments by adapting the H&M fashion recommendation dataset for experimentation. Our results show consistent and significant improvements when embedding is added as input to the ranking systems.