Location-Aware Named Entity Disambiguation

Abstract

Named Entity Disambiguation (NED) and linking has been traditionally evaluated on natural language content that is both well-written and contextually rich. However, many NED approaches display poor performance on text sources that are short and noisy. In this thesis, we study the problem of entity disambiguation for short text and propose a location-aware NED framework that resolves ambiguities in text with little other contextual cues. We show that the spatial dimension is crucial in disambiguating named entities and that the location inference is less utilized in many NED systems. Our proposed framework integrates (in an unsupervised manner) spatial signals that are readily available for many sources that emit short text (e.g., micro-blogs, search queries, and news streams). Our evaluation on news headlines and tweets reveals that a simple spatial embedding improves the accuracy of competitive baseline NED approaches from the literature by 8% for the news headlines and by 4% on tweets in probabilistic model. We further evaluated our spatial feature in a neural model and it showed that the NED performance is improved by 1.5% for the news headlines and by 6% for the tweets.