关键词: 方面-观点对抽取, 新能源汽车投诉文本, 语境增强, 三仿射机制, 多头注意力

Abstract: Mining users' multi-dimensional opinions on products in the complaint texts of new energy vehicles can provide support for product design decisions. Considering that the complaint text has the characteristics of high entity density and lengthy sentence structure, the existing methods for aspect-opinion pair extraction suffer from weak correlations between aspect terms and opinion terms. To locate this problem, this paper proposes an Aspect-Opinion pair Extraction model based on Context Enhancement(AOE-CE), fusing topic features and text features as contextual representation to enhance the correlation between entities. This model consists of an entity recognition module and relation detection module. Firstly, The entity recognition module encodes the text using a pre-trained model and part-of-speech tagging tool. Secondly, it employs Bi-directional Long Short-Term Memory (Bi-LSTM) combined with multi-head attention to capture contextual information and derive text features. Subsequently, These features are input into a Conditional Random Fields (CRF) model to obtain the entity set. The relationship detection module obtains the topic features through BERT, and fuses them with the text features to obtain the enhanced contextual representation. Then use the tri-affine mechanism to improve the relationship between entities with the help of contextual representation. Finally, the extraction result is obtained by Sigmoid. Meanwhile, the experimental result shows the accuracy rate, recognition rate and F1 value of the proposed method are 2.19%, 1.08% and 1.60% higher than the SDRN model respectively.

Key words: aspect-opinion pair extraction, the complaint texts of new energy vehicles, context enhancement, the tri-affine mechanism, multi-head attention