Embarrassingly Simple Unsupervised Aspect Extraction

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3182–3187 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3182 Embarrassingly Simple Unsupervised Aspect Extraction St´ephan Tulkens CLiPS University of Antwerp Belgium stephan.tulkens@uantwerpen.be Andreas van Cranenburgh Department of Information Science University of Groningen The Netherlands a.w.van.cranenburgh@rug.nl Abstract We present a simple but effective method for aspect identification in sentiment analysis. Our unsupervised method only requires word embeddings and a POS tagger, and is there- fore straightforward to apply to new domains and languages. We introduce Contrastive At- tention (CAt ), a novel single-head attention mechanism based on an RBF kernel, which gives a considerable boost in performance and makes the model interpretable. Previous work relied on syntactic features and complex neu- ral models. We show that given the sim- plicity of current benchmark datasets for as- pect extraction, such complex models are not needed. The code to reproduce the experi- ments reported in this paper is available at https://github.com/clips/cat. 1 Introduction We consider the task of unsupervised aspect ex- traction from text. In sentiment analysis, an as- pect can intuitively be defined as a dimension on which an entity is evaluated (see Figure 1). While aspects can be concrete (e.g., a laptop battery), they can also be subjective (e.g., the loudness of a motorcycle). Aspect extraction is an important subtask of aspect-based sentiment analysis. How- ever, most existing systems are supervised (for an overview, cf. Zhang et al., 2018). As aspects are domain-specific, supervised systems that rely on strictly lexical cues to differentiate between aspects are unlikely to transfer well between different do- mains (Rietzler et al., 2019). Another reason to con- sider the unsupervised extraction of aspect terms is the scarcity of training data for many domains (e.g., books), and, more importantly, the complete lack of training data for many languages. Unsupervised aspect extraction has previously been attempted with topic models (Mukherjee and Liu, 2012), topic model hybrids (Garc´ıa-Pablos et al., 2018), and re- The two things that really drew me to vinyl were the expense and the inconvenience . Figure 1: An example of a sentence expressing two aspects (red) on a target (italics). Source: https: //www.newyorker.com/cartoon/a19180 A: aspect vectors S: sentence (word vectors) att: attention vector d: sentence summary food staff ambience RBF Figure 2: An overview of our aspect extraction model. stricted Boltzmann machines (Wang et al., 2015), among others. Recently, autoencoders using atten- tion mechanisms (He et al., 2017; Luo et al., 2019) have also been proposed as a method for aspect extraction, and have reached state of the art perfor- mance on a variety of datasets. These models are unsupervised in the sense that they do not require labeled data, although they do rely on unlabeled data to learn relevant patterns. In addition, these are complex neural models with a large number of parameters. We show that a much simpler model suffices for this task. We present a simple unsupervised method for aspect extraction which only requires a POS …

Content truncated to 500 words