Effects on Time and Quality of Short Text Clustering during Real-Time Presentations

Technologies for live presentations should consider users' capabilities to manage large amounts of data in real-time, particularly, exchanges of short texts (e.g., phrases). This study examines the effects on time and quality of text clustering algorithms applied to short, medium, and long size texts, and examines whether short text clustering shows a reasonable performance for live presentations. We run several simulations in which we varied the number of phrases (from 5 to 200) contained in each text type (long, medium, and short) and the number of generated clusters (from 2 to 10). The algorithms used were snowball steamers, TF-IDF, and K-means for clustering; and the text types were Reuters, 20 NewsGroup and an experimental data set, for the long, medium, and short size texts, respectively. The first result showed that text size had a large effect on the algorithms execution time, with the shortest average time for the short texts and longer average time for the longest texts. The second result showed that the number of phrases in each text type significantly predicts execution time but not the number of clusters generated by K-means. Inertia and purity measures were used to test the quality of the clusters generated. Text size, number of phrases and number of clusters predict inertia; showing the lowest inertia for the short texts. Purity measures were like previously reported results for all text types. Thus, clustering algorithms for short texts can confidently be used in real-time presentations.