IHS R&D Belarus: Cross-domain extraction of product features using CRF

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 309–313, Dublin, Ireland, August 23-24, 2014. IHS R&D Belarus: Cross-domain Extraction of Product Features using Conditional Random Fields Maryna Chernyshevich IHS Inc. / IHS Global Belarus 131 Starovilenskaya St. 220123, Minsk, Belarus Marina.Chernyshevich@ihs.com Abstract This paper describes the aspect extraction system submitted by IHS R&D Belarus team at the SemEval-2014 shared task re- lated to Aspect-Based Sentiment Analy- sis. Our system is based on IHS Goldfire linguistic processor and uses a rich set of lexical, syntactic and statistical features in CRF model. We participated in two domain-specific tasks – restaurants and laptops – with the same system trained on a mixed corpus of reviews. Among sub- missions of constrained systems from 28 teams, our submission was ranked first in laptop domain and fourth in restaurant domain for the subtask A devoted to as- pect extraction. 1 Introduction With a rapid growth of the blogs, forums, review sites and social networks, more and more people express their personal views about products on the Internet in form of reviews, ratings, or rec- ommendations. This is a great source of data used by many researchers and commercial appli- cations that are focused on the sentiment analy- sis to determine customer opinions. Sentiment analysis can be done on document, sentence, and phrase level (Jagtap, V. S., Ka- rishma Pawar, 2013). Earlier works were focused mainly on the document (Turney, 2002; Pang, Lee and Vaithyanathan, 2002) and the sentence level (Kim and Hovy, 2004). However, this in- formation can be insufficient for customers who are seeking opinions on specific product features (aspects) such as design, battery life, or screen. This fine-grained classification is a topic of as- This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons. org/licenses/by/4.0/ pect-based sentiment analysis (Moghaddam and Ester, 2012). Traditional approaches to aspect extraction are based on frequently used nouns and noun phrases (Popescu and Etzioni, 2005; Blair-Goldensohn et al., 2008), exploiting opinions (Zhuang et al., 2006; Kobayashi, 2006), and supervised learning (Mukherjee and Liu, 2012). In this paper, we describe a system (IHS_RD_Belarus in official results) developed to participate in the international shared task or- ganized by the Conference on Semantic Evalua- tion Exercises (SemEval-2014) and focused on the phrase-level sentiment classification, namely aspect extraction (Pontiki et al., 2014). An aspect term means particular feature of a product or ser- vice used in opinion-bearing sentences (My phone has amazing screen), as well as in neutral sentences (The screen brightness automatically adjusts). The organizers of SemEval-2014 task have provided a dataset of customer reviews with an- notated aspects of the target entities from two domains: restaurants (3041 sentences) and lap- tops (3045 sentences). The results were evaluat- ed separately in each domain. Table 1 shows the distribution of the provided data for each domain dataset, training and testing set, with number of sentences and aspects. Laptops Restaurants Training Sentences 3045 …

Content truncated to 500 words