Entity-Linking via Graph-Distance Minimization

TL;DR

This paper introduces a new graph-based approach to entity-linking, focusing on minimizing graph-distance to improve wikification, and proposes heuristics for NP-hard instances with experimental validation.

Contribution

It formalizes a novel graph problem related to entity-linking, proves its NP-hardness, and offers practical heuristics with empirical evaluation.

Findings

Heuristics perform well compared to baselines

Problem is NP-hard in general but solvable under certain conditions

Proposed methods improve entity-linking accuracy

Abstract

Entity-linking is a natural-language-processing task that consists in identifying the entities mentioned in a piece of text, linking each to an appropriate item in some knowledge base; when the knowledge base is Wikipedia, the problem comes to be known as wikification (in this case, items are wikipedia articles). One instance of entity-linking can be formalized as an optimization problem on the underlying concept graph, where the quantity to be optimized is the average distance between chosen items. Inspired by this application, we define a new graph problem which is a natural variant of the Maximum Capacity Representative Set. We prove that our problem is NP-hard for general graphs; nonetheless, under some restrictive assumptions, it turns out to be solvable in linear time. For the general case, we propose two heuristics: one tries to enforce the above assumptions and another one is based on the notion of hitting distance; we show experimentally how these approaches perform with respect to some baselines on a real-world dataset.