Extracting Descriptions of Location Relations from Implicit Textual Networks

Abstract

For the retrieval of concise entity relation information from large collections or streams of documents, existing approaches can be grouped into the categories of (multi-document) summarization and knowledge extraction. The former tend to fall short for this task due to the involved amount of information that cannot be easily condensed, while knowledge extraction approaches are often pattern-based and too discriminative for exploratory purposes. For location relations in particular, this translates to a set of very short relationship descriptors that predominantly encode hierarchical or containment relations such as located in or capital of. As a result, available knowledge bases that are typically populated through knowledge extraction are limited to these discrete and typed relations. In contrast, the representation of document collections as implicit networks of entities, terms, and sentences has emerged as a way to encode a much wider range of entity relations and occurrences, which can be leveraged for filtering the relevant information and enabling subsequent interactive explorations. In this paper, we discuss the extraction of descriptive sentences for sets of entities from such implicit networks to support an interactive exploration, and apply them to the extraction of complex location relations that are not hierarchical or containment-based. We introduce and compare efficient ranking methods for sentence extraction that address this entity-centric search task by leveraging entity and term relations in implicit network representations of large document collections. Based on Wikipedia articles and Wikidata as a knowledge base, we demonstrate the extraction of novel location relations that are not contained in the knowledge base.