Brand-Topic Model (BTM) is a model aims to detect brand-associated polarity-bearing topics from product reviews. Different from existing models for sentiment-topic extraction which assume topics are grouped under discrete sentiment categories such as `positive', `negative' and `neural', BTM is able to automatically infer real-valued brand-associated sentiment scores and generate fine-grained sentiment-topics in which we can observe continuous changes of words under a certain topic (e.g., `shaver' or `cream') while its associated sentiment gradually varies from negative to positive. Example topics generated by BTM are shown below.
We construct our dataset by retrieving reviews in the Beauty category from the Amazon review corpus. Each review is accompanied with the rating score, reviewer name and the product meta-data such as product ID, description, brand and image.
Since consumer purchase decisions are heavily influenced by online reviews, it is important to automatically analyze customer reviews for on-line brand monitoring. As output we aim at identifying the fine-grained topics that are shared by multiple brands and generate fine-grained brand score for each brand, from -1 to 1.
BTM is built on the Poisson Factorisation Model with the incorporation of adversarial learning. Our model is partly inspired by TBIP, which is used for modelling idea points in political debating.
First the model is pre-trained with the Poisson Factorisation Model, then we initialise TBIP model with the doc-topics matrix and topic-word matrix inherited from the pre-training. In order to adjust the class imbalance in the training data (#positive reviews >> #negative reviews), we propose adversarial learning by negating the brand-polarity score. With the reversed Poisson distribution, we apply Gumbel-Softmax to sample a reversed review from the distribution. Finally the reversed review was fed to a sentiment classifier to generate a rating score.
The overall architecture of BTM is shown in the following figure.
R. Zhao, L. Gui, G. Pergola and Y. He. Adversarial Learning of Poisson Factorisation Model for Gauging Brand Sentiment in User Reviews, The 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Apr. 2021.