Coupling Global and Local Context for Unsupervised Aspect Extraction

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4579–4589, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4579 Coupling Global and Local Context for Unsupervised Aspect Extraction Ming Liao1,2∗, Jing Li3,†, Haisong Zhang4, Lingzhi Wang1,2, Xixin Wu1, Kam-Fai Wong1,2 1The Chinese University of Hong Kong, Hong Kong, China 2MOE Laboratory of High Confidence Software Technologies, China 3Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China 4Tencent AI Lab, Shenzhen, China 1,2{mliao,lzwang,wuxx,kfwong}@se.cuhk.edu.hk 3jing-amilia.li@polyu.edu.hk, 4hansonzhang@tencent.com Abstract Aspect words, indicating opinion targets, are essential in expressing and understanding hu- man opinions. To identify aspects, most pre- vious efforts focus on using sequence tag- ging models trained on human-annotated data. This work studies unsupervised aspect extrac- tion and explores how words appear in global context (on sentence level) and local context (conveyed by neighboring words). We pro- pose a novel neural model, capable of cou- pling global and local representation to dis- cover aspect words. Experimental results on two benchmarks, laptop and restaurant re- views, show that our model significantly out- performs the state-of-the-art models from pre- vious studies evaluated with varying metrics. Analysis on model output show our ability to learn meaningful and coherent aspect repre- sentations. We further investigate how words distribute in global and local context, and find that aspect and non-aspect words do exhibit different context, interpreting our superiority in unsupervised aspect extraction. 1 Introduction Opinion, one of the main factors shaping human behavior, is crucial to our daily activities (Liu, 2012). Every choice we make in our life, rang- ing from where to go for a Friday dinner to which job offer to pick up, is largely influenced by what other people think. To help individuals navigate decision-making processes, there exists growing attentions on opinion mining algorithms that dis- till massive opinion-rich texts — such as digital product reviews (Poddar et al., 2017) and social media discussions (Dusmanu et al., 2017) — into the opinionated information we need. ∗ This work was partially done when Ming Liao was an intern at Tencent AI Lab, Shenzhen, China. †Jing Li is the corresponding author and conducted most of the work at Tencent AI Lab, Shenzhen, China. [R1]: It’s truly a great laptop::: for :: the::::: price. [R2]: If you have the money, I suggest going :: for :: the::i7 Table 1: Two sample laptop review sentences. Aspect words are in boldface and blue. Wavy underlines indi- cate local context words indicating aspect word “i7”. Towards human opinion understanding, it is es- sential to figure out what target the opinion cen- ters around. After all, previous studies have long pointed out that human language mostly conveys opinion with aspect and sentiment words (Liu, 2012). In this work, we focus on aspect extrac- tion, targeting at the recognition of words indi- cating opinion aspects (henceforth aspect words). We believe developing effective aspect extraction models will benefit a broad range of compelling applications, such …

Content truncated to 500 words