SemEval-2015 Task 12: Aspect Based Sentiment Analysis

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 486–495, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics SemEval-2015 Task 12: Aspect Based Sentiment Analysis Abstract SemEval-2015 Task 12, a continuation of SemEval-2014 Task 4, aimed to foster re- search beyond sentence- or text-level senti- ment classification towards Aspect Based Sentiment Analysis. The goal is to identify opinions expressed about specific entities (e.g., laptops) and their aspects (e.g., price). The task provided manually annotated reviews in three domains (restaurants, laptops and ho- tels), and a common evaluation procedure. It attracted 93 submissions from 16 teams. 1 Introduction and Related Work The rise of e-commerce, as a new shopping and marketing channel, has led to an upsurge of review sites for a variety of services and products. In this context, Aspect Based Sentiment Analysis (ABSA) -i.e., mining opinions from text about specific enti- ties and their aspects- can help consumers decide what to purchase and businesses to better monitor their reputation and understand the needs of the market (Pavlopoulos 2014). Given a target of in- terest (e.g., Apple Mac mini), an ABSA method can summarize the content of the respective re- views in an aspect-sentiment table like the one in Fig 1. Some review sites also generate such tables based on customer ratings, but usually only for a limited set of predefined aspects and not from free- text reviews. Several ABSA methods have been proposed for various domains, like consumer electronics (Hu and Liu {2004a, 2004b}), restaurants (Ganu et al., 2009) and movies (Thet et al., 2010). The available methods can be divided into those that adopt do- main-independent solutions (Lin and He, 2009), and those that use domain-specific knowledge to improve their results (Thet et al., 2010). Typically, most methods treat aspect extraction and sentiment classification separately (Brody and Elhadad, 2010), but there are also approaches that model the two problems jointly (Jo and Oh, 2011). Figure 1. Table summarizing the average sentiment for each aspect of an entity. Publicly available ABSA datasets adopt differ- ent annotation schemes for different subtasks and languages (Pavlopoulos 2014). For example, the datasets of McAuley et al. (2012) provide aspects and respective ratings at the review level (i.e., as- pects and ratings associated with entire reviews, not particular sentences)1 about Beers, Pubs, Toys and Games, and Audiobooks. The reviews are ob- tained from sites that allow users to evaluate a product not only in terms of its overall quality, but also focusing on specific predefined aspects (e.g. “smell” and “taste” for Beers, “fun” and “educa- tional value” for Toys and Games). The IGGSA Shared Tasks on German Sentiment Analysis (Ruppenhofer et al., 2014) provided human anno- tated datasets of political speeches (STEPS task) 1 A subset of the datasets has been annotated with aspects at the sentence level. Maria Pontiki*, Dimitrios Galanis*, Haris Papageorgiou*, Suresh Manandhar±, Ion Androutsopoulos◊* *Institute for Language and Speech Processing, Athena R.C., Athens, Greece ± Dept. of Computer Science, University of York, UK ◊ Dept. …

Content truncated to 500 words