Knowledge graph based entity selection framework for ad-hoc retrieval

Recent entity-based retrieval models utilizing knowledge bases have shown significant improvement in ad-hoc retrieval. However, a lack of coherence between candidate entities can lead to query intent drift at retrieval time. To address this issue, we present an entity selection algorithm that utilizes a graph clustering framework to discover the semantics between entities and encompass the query with highly coherent entities accumulated from different resources, including knowledge bases, and pseudo-relevance feedback documents. Through this work, we propose: (1) An entity acquisition strategy to systematically acquire coherent entities for query expansion. (2) We propose a graph representation of entities to capture the coherence between entities where nodes correspond to the entities and edges represent semantic relatedness between entities. (3) We propose two different entity ranking approaches to select candidate entities based on the coherence with query entities and other coherent entities. A set of experiments on five TREC collections: ClueWeb09B, ClueWeb12B, Robust04, GOV2, and MS-Marco dataset under document retrieval task were conducted to verify the proposed algorithm’s performance. The reported results indicated that the proposed methodology outperforms existing state-of-the-art retrieval approaches in terms of MAP, NDCG, and P@20. The code and relevant data are available in https://github.com/pankajkashyap65/KnowledgeGraph.