Abstract
The application of topic modeling to short texts is beset by challenges such as data sparsity and an absence of contextual information. Traditional research methods tend to prioritise high-attention and popular topics, frequently overlooking the identification of emerging topics. Consequently, subjects of a minor scale are prone to being overlooked during the topic identification process. Furthermore, in the context of topic modelling, information that varies in terms of the attention it receives is not treated equally. In order to address the aforementioned issues, a fairness-oriented topic discovery approach (MixTM-G) is proposed. This approach has been designed to facilitate the discovery of topics with different levels of attention. The proposed methodology involves the integration of normalized pointwise mutual information (NPMI) within a graph model to analyse text data. This approach leverages the correlation between data points to assess the semantic relationships between words, thus addressing the limitations posed by sparse data. The employment of graph algorithms facilitates the identification of semantically related clusters within the document graph, thereby enhancing the semantic associations between sparse data. Finally, a mixed topic modeling approach (MixTM), based on bi-grams and tri-grams combinations, is proposed to further improve topic discovery by strengthening the contextual relationships between words. The experimental results demonstrate the efficacy of the proposed method in topic modelling. In comparison to conventional methods, the proposed approach exhibits superior performance in detecting small-scale topics under equivalent conditions.