H-ERNIE: A Multi-Granularity Pre-Trained Language Model for Web Search

The pre-trained language models (PLMs), such as BERT and ERNIE, have achieved outstanding performance in many natural language understanding tasks. Recently, PLMs-based Information Retrieval models have also been investigated and showed substantially state-of-the-art effectiveness, e.g., MORES, PROP and Co1BERT. Moreover, most of the PLMs-based rankers only focus on a single level relevance matching (e.g., character-level), while ignore the other granularity information (e.g., words and phrases), which easily lead to the ambiguity of query understanding and inaccurate matching issues in web search. In this paper, we aim to improve the state-of-the-art PLMs ERNIE for web search, by modeling multi-granularity context information with the awareness of word importance in queries and documents. In particular, we propose a novel H-ERNIE framework, which includes a query-document analysis component and a hierarchical ranking component. The query-document analysis component has several individual modules which generate the necessary variables, such as word segmentation, word importance analysis, and word tightness analysis. Based on these variables, the importance-aware multiple-level correspondences are sent to the ranking model. The hierarchical ranking model includes a multi-layer transformer module to learn the character-level representations, a word-level matching module, and a phrase-level matching module with word importance. Each of these modules models the query and the document matching from a different perspective. Also, these levels are inherently communicated to achieve the overall accurate matching. We discuss the time complexity of the proposed framework, and show that it can be efficiently implemented in real applications. The offline and online experiments on both public datasets and a commercial search engine illustrate the effectiveness of the proposed H-ERNIE framework.