Robustness and Sensitivity of Network-Based Topic Detection

Abstract

In the context of textual analysis, network-based procedures for topic detection are gaining attention as an alternative to classical topic models. Network-based procedures are based on the idea that documents can be represented as word co-occurrence networks, where topics are defined as groups of strongly connected words. Although many works have used network-based procedures for topic detection, there is a lack of systematic analysis of how different design choices, such as the building of the word co-occurrence matrix and the selection of the community detection algorithm, affect the final results in terms of detected topics. In this work, we present the results obtained by analysing a widely used corpus of news articles, showing how and to what extent the choices made during the design phase affect the results.