TRIO: An Entity Retrieval Method Using Entity Embedding and Topic Modeling

Hyejin Park, Daecheol Woo, Shinwoo Park, Kyungwon Kim

Abstract: Current entity search models predominantly rely on term frequency and semantic similarity, often failing to fully exploit the information in the knowledge graphs. This limitation leads to the neglect of entities that could be highly relevant to the user’s query topics. To overcome these challenges and enhance entity retrieval, we introduce TRIO (Term, topic, and neural-based entity Retrieval Interpolate methOd), an entity retrieval method that employs multiple search perspectives for more relevant outcomes. TRIO stands out by seamlessly integrating three distinct search perspectives: term frequency, semantic similarity, and topic similarity. This integration is executed in a simple and effective manner, allowing TRIO to capture entities across multiple dimensions, resulting in comprehensive and accurate search results. Our experiments on the standard DBpedia-Entity V2 test collection demonstrate a substantial enhancement in the search performance of the baseline model. On average, TRIO improves NDCG and MAP performance by 12.401%, and 27.342%, respectively, compared to the best-performing baseline model.