Topic Coherence Metrics: How Sensitive Are They?

Topic modeling approaches extract the most relevant sets of words (grouped into so-called topics) from a document collection. The extracted topics can be used for analyzing the latent semantic structure hiding in the collection. This task is intrinsically unsupervised (without information about the labels), so evaluating the quality of the discovered topics is challenging. To address that, different unsupervised metrics have been proposed, and some of them are close to human perception, e.g., coherence metrics. Moreover, metrics behave differently when facing noise (i.e., unrelated words) in the topics. This article presents an exploratory analysis to evaluate how state-of-the-art metrics are affected by perturbations in the topics. By perturbation, we mean that intruder words are synthetically inserted into the topics to measure the metrics’ ability to deal with noises. Our findings highlight the importance of overlooked choices in the metrics sensitiveness context. We show that some topic modeling metrics are highly sensitive to disturbing; others can handle noisy topics with minimal perturbation. As a result, we rank the chosen metrics by sensitiveness, and as the contribution, we believe that the results might be helpful for developers to evaluate the discovered topics better.