Term-Based Models for Entity Ranking

Abstract

We have established in Chap. 1 that entities are natural and meaningful units of retrieval. To recap, according to our working definition, an entity is a uniquely identifiable "thing," a typed object, with name(s), attributes, and relationships to other entities. Examples of some of the most frequent types of entities include people, locations, organizations, products, and events. Returning specific entities, instead of a mere list of documents, can provide better answers to a broad range of information needs. The way these information needs are expressed may vary from short keyword queries to full-fledged natural language questions (cf. Sect. 1.3.1.1). In this chapter, we adhere to the "single search box" paradigm, which accepts "free text" search queries, and simply treat queries as sequences of words, referred to hereinafter as terms. The task we are going to focus on is ad hoc entity retrieval: answering a one-off free text query, representing the user's underlying information need, with a ranked list of entities. The fundamental question we are concerned with then is: How do we perform relevance matching of queries against entities?In Chap. 2, we have introduced large-scale knowledge repositories, like Wikipedia and DBpedia, that are devoted to organizing information around entities in a (semi-)structured format. Having information accumulated about entities in a knowledge repository is indeed helpful, yet it is not a prerequisite. As we shall see later in this chapter, it is possible to rank entities without (pre-existing) direct representations, as long as they can be recognized and identified uniquely in documents. The main idea of this chapter can be summarized as follows: If textual representations can be constructed for entities, then the ranking of these representations ("entity descriptions") becomes straightforward by building on traditional document retrieval techniques (such as language models or BM25). Accordingly, the bulk of our efforts in this chapter, described in Sect. 3.2, will revolve around assembling term-based entity representations from various sources, ranging from unstructured to structured. In the simplest case, these entity representations are based on the bag-of-words model, which considers the frequency of words but disregards their order. We will also introduce an extension to multiple document fields, as a mechanism for preserving some of the structure associated with entities.