A Method to Estimate Entity Performance from Mentions to Related Entities in Texts on the Web

Abstract

Publications on the Web can influence the public opinion about certain entities (e.g., politicians, institutions). At the same time, a variety of indicators can be extracted from these publications and used to estimate entity performance (e.g., popularity, votes share). This work proposes an automatic method that employs state-of-the-art natural language processing tools to extract indicators about entities mentioned in texts, for estimating the performance of these entities or semantically related ones. Our method calculates performance metrics from performance indicators consolidated for semantically related entities, assess correlations of these consolidated metrics with ground true performance, and uses these metrics to predict certain fluctuations in entity performance. Experimental results in a case study on politics show that consolidated metrics for several interrelated entities are better correlated to observed real performance measures of some target entities and lead to better predictions, than metrics for just one entity.