Mixterm Topic Model: A Short Text Topic Model in Low-Resource Scenarios
Chuangying Zhu, Yongyu Liang, Limiao Zhong, Xinyuan Liang, Fei Xie · 2025
In online platforms, emerging topics often receive insufficient attention, resulting in a limited amount of related data. Traditional topic modeling methods rely on a large volumes of data, making it difficult to identify new topics in a timely manner and effectively track the topic of small-scale data. To address this issue, we propose a topic modeling method for small-scale data, namely MixTM. The model combines three words into a semantic unit to model short text data, thereby constraining term associations and expanding contextual information to provide more effective insights for small-scale topic identification. The proposed method alleviates the bias of traditional modeling approaches towards mature topics and offers technical support for tracking topics of varying scales in large datasets. Experimental results indicate that the proposed model outperforms traditional methods while successfully identifies small-scale topics even in datasets with limited volume.