Enhancing Heterogeneous Graph-based Short Text Topic Learning
Qingren Wang, Junwei Wu, Jie Cui · 2022
Short texts generated on social networks have become a widespread format of information. Multiple applications require a semantic understanding of text content, so deriving coherent topics from short texts is vital. However, traditional topic models lack sufficient word co-occurrence information limited by the sparsity of short text. The development of word embedding technology facilitates us to use the semantic and syntactic information in words for short text research. Word vectors consider the characteristics of words in various contexts when calculating the correlation among words, which is appropriate for short text topic learning. Therefore, this paper proposes a novel topic learning method called E-ShoTT, which combines Heterogeneous Graphs with prior semantic knowledge derived from word embeddings. E-ShoTT guides the extraction of phrase instances by constructing a heterogeneous graph regarding parts of speech while increasing the probability that semantically related words share the same topic assignment during topic sampling. Experiment results on two public benchmark datasets show E-ShoTT outperforms other methods in terms of topic coherence, cluster and classification accuracy.