SpanMlt: A Span-based Multi-Task Learning Framework for Pair-wise Aspect and Opinion Terms Extraction

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3239–3248 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3239 SpanMlt: A Span-based Multi-Task Learning Framework for Pair-wise Aspect and Opinion Terms Extraction He Zhao ∗, Longtao Huang ∗†, Rong Zhang, Quan Lu, Hui Xue Alibaba Group {sicheng.zh,kaiyang.hlt,stone.zhangr,luquan.lq,hui.xueh}@alibaba-inc.com Abstract Aspect terms extraction and opinion terms ex- traction are two key problems of fine-grained Aspect Based Sentiment Analysis (ABSA). The aspect-opinion pairs can provide a global profile about a product or service for con- sumers and opinion mining systems. How- ever, traditional methods can not directly out- put aspect-opinion pairs without given aspect terms or opinion terms. Although some recent co-extraction methods have been proposed to extract both terms jointly, they fail to extract them as pairs. To this end, this paper pro- poses an end-to-end method to solve the task of Pair-wise Aspect and Opinion Terms Extrac- tion (PAOTE). Furthermore, this paper treats the problem from a perspective of joint term and relation extraction rather than under the se- quence tagging formulation performed in most prior works. We propose a multi-task learn- ing framework based on shared spans, where the terms are extracted under the supervision of span boundaries. Meanwhile, the pair-wise relations are jointly identified using the span representations. Extensive experiments show that our model consistently outperforms state- of-the-art methods. 1 Introduction Fine-grained aspect-based sentiment analysis (ABSA) or opinion mining is a field of study that analyzes people’s detailed insights towards a prod- uct or service. Aspect terms (AT) extraction and opinion terms (OT) extraction are two fundamen- tal subtasks in ABSA (Pang and Lee., 2008; Liu, 2012). Aspect terms, also named as opinion targets, are the word sequences in the sentence describing attributes or features of the targets. Opinion terms, sometimes called opinion words, are those expres- sions carrying subjective attitudes. For example, ∗Both authors contributed equally to this research. †Corresponding author Figure 1: An example of the difference between co- extraction and pair extraction of AT and OT. in the sentence “Otherwise, this place has great service and prices and a nice friendly atmosphere”, the aspect terms are service, prices and atmosphere, and the opinion terms are great and nice friendly. Recently, a new research focus, which aims at co-extracting the aspect and opinion terms (Wang et al., 2016, 2017; Li and Lam, 2017; Wang and Pan, 2018; Yu et al., 2019), has drawn increas- ing attention in both academia and industry. Such methods use joint models and have achieved great progress on both subtasks. However, the extracted AT and OT are not in pairs, and the correspond- ing relations between them are not well extracted. As the example sentence shown in Figure 1, (ser- vice, great), (prices, great) and (atmosphere, nice friendly) are three aspect-opinion pairs. In contrast, the co-extraction methods can only output the AT set {service, prices, atmosphere} and the OT set {great, nice friendly} jointly. The aspect-opinion pairs can deploy more fine- grained sentiment analysis for review text …

Content truncated to 500 words