Short Text Tagging Using Nested Stochastic Block Model: A Yelp Case Study

Abstract

From online reviews and product descriptions to tweets and chats, many modern applications revolve around understanding both semantic structure and topics of short texts. Due to significant reliance on word co-occurrence, traditional topic modeling algorithms such as LDA perform poorly on sparse short texts. In this paper, we propose an unsupervised short text tagging algorithm that generates latent topics, or clusters of semantically similar words, from a corpus of short texts, and labels these short texts by stable predominant topics. The algorithm defines a weighted undirected network, namely the one mode projection of the bipartite network between words and users. Nodes represent all unique words from the corpus of short texts, edges mutual presence of pairs of words in a short text, and weights the number of short texts in which pairs of words appear. We generate the latent topics using nested stochastic block models (NSBM), dividing the network of words into communities of similar words. The algorithm is versatile—it automatically detects the appropriate number of topics. Many applications stem from the proposed algorithm, such as using the short text topic representations as the basis of a short text similarity metric. We validate the results using inter-semantic similarity and normalized mutual information, which show the method is competitive with industry short text topic modeling algorithms.