Aspect-Based Pair-Wise Opinion Generation in Chinese automotive reviews: Design of the task, dataset and model

Abstract

The extraction of opinion target–word pairs from user reviews has received much attention recently, since it can provide essential information for fine-grained opinion analysis. However, according to our statistics on a large-scale dataset of Chinese reviews, about 60% reviews do not explicitly show opinion targets or words. To investigate this problem, this paper introduces a new task under fine-grained opinion analysis, named Aspect-Based Pair-wise Opinion Generation (ABPOG), which aims to generate opinion target–word pairs based on reviews and aspects. To perform this task, we develop a sequence-to-sequence model for opinion target–word pair generation by extending the pointer–generator network with two approaches: (1) an aspect-aware encoder that receives an additional aspect embedding as input to extract aspect-specific features, (2) two hierarchical decoders including a token-level GRU and a global GRU to generate opinion targets and words jointly. To empirically evaluate our task and model, we develop a multi-aspect dataset for ABPOG based on Chinese automotive reviews. Extensive experiments on our dataset show that our model outperforms several strong baselines adapted from the state-of-the-art aspect-based summarization method.