TRENDNERT: A Benchmark for Trend and Downtrend Detection in a Scientific Domain

Abstract

Computational analysis and modeling of the evolution of trends is an important area of research in Natural Language Processing (NLP) because of its socio-economic impact. However, no large publicly available benchmark for trend detection currently exists, making a comparative evaluation of methods impossible. We remedy this situation by publishing the benchmark TRENDNERT, consisting of a set of gold trends and downtrends and document labels that is available as an unrestricted download, and a large underlying document collection that can also be obtained for free. We propose Mean Average Precision (MAP) as an evaluation measure for trend detection and apply this measure in an investigation of several baselines.