A Multifacet Hierarchical Sentiment-Topic Model with Application to Multi-Brand Online Review Analysis

TL;DR

This paper tackles automatic multi-brand comparison from online reviews by modeling hierarchical, shared aspects and brand-specific polarity across brands. To do so, it introduces MH-STM, a multifacet hierarchical sentiment-topic model that uses a tree-structured topic space with depth $L$ and per-brand $nCRP$ to allocate sentences to paths, plus a regression on empirical topic distributions to predict review polarity. It couples this with a hierarchical Pólya urn (HPU) to weight topic-word associations across the hierarchy via entropy-based addition weights, improving separation of general root topics from brand-specific leaf topics. Across synthetic data and two real corpora, MH-STM yields higher topic coherence and more accurate multi-aspect brand rankings, while SHLDA demonstrates better language fit but less interpretable hierarchies. These results suggest a practical, extensible framework for competitive benchmarking and marketing insights derived from user reviews.

Abstract

Multi-brand analysis based on review comments and ratings is a commonly used strategy to compare different brands in marketing. It can help consumers make more informed decisions and help marketers understand their brand's position in the market. In this work, we propose a multifacet hierarchical sentiment-topic model (MH-STM) to detect brand-associated sentiment polarities towards multiple comparative aspects from online customer reviews. The proposed method is built on a unified generative framework that explains review words with a hierarchical brand-associated topic model and the overall polarity score with a regression model on the empirical topic distribution. Moreover, a novel hierarchical Polya urn (HPU) scheme is proposed to enhance the topic-word association among topic hierarchy, such that the general topics shared by all brands are separated effectively from the unique topics specific to individual brands. The performance of the proposed method is evaluated on both synthetic data and two real-world review corpora. Experimental studies demonstrate that the proposed method can be effective in detecting reasonable topic hierarchy and deriving accurate brand-associated rankings on multi-aspects.

Figures (6)

• Figure 1: Illustration of laptop reviews with respect to various brands. Each review is composed of a plain text and an overall rating. Different brands share multiple comparative aspects represented by a tree structure of product aspect nodes with respective distributions on that, where each aspect node is linked to multiple brand-specific sentiment polarities that influence the overall ratings in brand reviews.
• Figure 2: Graphical representation of MH-STM. The shaded circles represent observable variables, dashed circles represent hyperparameters, dotted circles represent latent variables, and the rest solid circles represent model parameters.
• Figure 3: Comparison of simple and hierarchical Pólya urn schemes in hierarchical topic modeling. Urns represent topics and balls with different numbers/colors represent different words. Grey balls (No.1-2) represent general words that are shared among child topics, while balls in color of blue (No.3-5), green (No.6-8), and orange (No.9-11) are specialized words for each child topic, respectively.
• Figure 4: Topic trees inferred from synthetic data. Each cell corresponds to a single word, and is shaded with intensity proportional to the probability of each word in the topic.
• Figure 5: Portions of 3-level topic hierarchy discovered from laptop reviews by SHLDA and MH-STM. Each topic is represented by its top 10 words with corresponding word probabilities under the topic. The numbers on the right side of each topic are brand-specific regression parameters learned with that topic.