# Tracking Brand-Associated Polarity-Bearing Topics in User Reviews

**Authors:** Runcong Zhao, Lin Gui, Hanqi Yan, Yulan He  
**arXiv:** [2301.07183](https://arxiv.org/abs/2301.07183)  
**Date:** 2023-01-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.07183v1/

## Key points

- Importance of monitoring online customer reviews for businesses
- Proposal of a dynamic Brand-Topic Model (dBTM) to detect and track brand-associated sentiment scores and polarity-bearing topics in product reviews
- Use of Gaussian state space models to capture the evolution of brand polarity scores and topic-word distributions over time
- Incorporation of a meta learning strategy for smooth topic transitions and enhanced brand score predictions
- Evaluation using MakeupAlley reviews and a hotel review dataset, demonstrating superior performance compared to baselines in terms of brand ranking
- Achievement of favorable balance between topic coherence and uniqueness, extracting well-separated polarity-bearing topics across different time intervals
- Significance of dBTM in offering an automated approach for tracking brand-associated sentiment scores and polarity-bearing topics over time
- Accurate brand score predictions with smooth topic transitions due to the use of Gaussian state space models and meta learning strategy
- Validation of dBTM's superiority over existing baselines in terms of brand ranking, topic coherence, uniqueness, and temporal separation of polarity-bearing topics.

## Summary

The paper titled "Tracking Brand-Associated Polarity-Bearing Topics in User Reviews" emphasizes the importance of monitoring online customer reviews for businesses to assess customer satisfaction and effectively manage their reputations. To address this need, the authors propose a novel dynamic Brand-Topic Model (dBTM) that can automatically detect and track brand-associated sentiment scores and polarity-bearing topics from product reviews organized in temporally-ordered time intervals. The dBTM model is designed to capture the evolution of latent brand polarity scores and topic-word distributions over time using Gaussian state space models. Additionally, it incorporates a meta learning strategy to control the update of the topic-word distribution in each time interval, ensuring smooth topic transitions and enhancing brand score predictions. To evaluate the effectiveness of dBTM, the authors utilize datasets constructed from MakeupAlley reviews and a hotel review dataset. The experimental results demonstrate that dBTM outperforms several competitive baselines in terms of brand ranking. It achieves a favorable balance between topic coherence and uniqueness while successfully extracting well-separated polarity-bearing topics across different time intervals. Overall, this research highlights the significance of monitoring online customer reviews for businesses. The proposed dBTM model offers an automated approach to detect and track brand-associated sentiment scores and polarity-bearing topics over time. By leveraging Gaussian state space models and incorporating a meta learning strategy, dBTM provides accurate brand score predictions with smooth topic transitions. The evaluation on real-world datasets further validates its superiority over existing baselines in terms of brand ranking, topic coherence, uniqueness, and temporal separation of polarity-bearing topics.