Multi-label Aspect Classification on Question-Answering Text with Contextualized Attention-Based Neural Network

Abstract

In the e-commerce websites, such as Taobao and Amazon, interactive question-answering (QA) style reviews usually carry rich aspect information of products. To well automatically analyze the aspect information inside QA style reviews, it’s worthwhile to perform aspect classification on them. Unfortunately, until now, there are few papers that focus on performing aspect classification on the QA style reviews. For short, we referred to this novel task as QA aspect classification (QA-AC). In this study, we model this task as a multi-label classification problem where each QA style review is explicitly mapped to multiple aspect categories instead of only one aspect category. To solve this issue, we propose a contextualized attention-based neural network approach to capture both the contextual information and the QA matching information inside QA style reviews for the task of QA-AC. Specifically, we first propose two aggregating strategies to integrate multi-layer contextualized word embeddings of the pre-trained language representation model (i.e., BERT) so as to capture contextual information. Second, we propose a bidirectional attention layer to capture the QA matching information. Experimental results demonstrate the effectiveness of our approach to QA-AC.