A clustering-based topic model using word networks and word embeddings

Abstract Online social networking services like Twitter are frequently used for discussions on numerous topics of interest, which range from mainstream and popular (e.g., music movies) to niche specialized politics). Due the popularity such services, it is a challenging task automatically model determine discussion given large amount tweets. Adding this complexity need identify these with absence prior knowledge about both types number topics, while having requirement relevant technical expertise tune parameters various models. To address challenge, we develop Clustering-based Topic Modelling (ClusTop) algorithm that first constructs different word networks based n-grams co-occurrence embedding distances. Using networks, ClusTop then able using community detection approaches. In contrast traditional topic models, does not require tuning or setting instead uses approaches appropriate topics. The also capture syntactic meaning in tweets via use bigrams, trigrams, other combinations techniques constructing network graph, utilizes edge weights embedding. three datasets labelled crises events as show outperforms baselines terms coherence, pointwise mutual information, precision, recall F-score.