Entity retrieval via type taxonomy aware smoothing

Abstract

We investigate the task of ad-hoc entity retrieval from a knowledge base. We propose a type taxonomy aware smoothing method that exploits the hierarchical type information of a knowledge base and integrates into an existing language modelling framework. Unlike most existing type-aware retrieval models, our approach does not require an explicit inference of query type. Instead, it directly encodes the type information into a term-based retrieval model by considering the occurrence of query terms in multi-fielded pseudo documents of entities whose types have connections in the type taxonomy. We conduct experiments on a recent public benchmark dataset with the Wikipedia category information. Preliminary experiment results show that our framework improves the performance of existing models.