USF: Chunking for Aspect-term Identification & Polarity Classification

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 790–795, Dublin, Ireland, August 23-24, 2014. USF: Chunking for Aspect Term Identification & Polarity Classification Cindi Thompson University of San Francisco 2130 Fulton St, HR 240 San Francisco, CA 94117 cathompson4@usfca.edu Abstract This paper describes the systems submit- ted by the University of San Francisco (USF) to Semeval-2014 Task 4, Aspect Based Sentiment Analysis (ABSA), which provides labeled data in two domains, lap- tops and restaurants. For the constrained condition of both the aspect term extrac- tion and aspect term polarity tasks, we take a supervised machine learning approach using a combination of lexical, syntactic, and baseline sentiment features. Our ex- traction approach is inspired by a chunk- ing approach, based on its strong past re- sults on related tasks. Our system per- formed slightly below average compared to other submissions, possibly because we use a simpler classification model than prior work. Our polarity labeling ap- proach uses two baseline hand-built sen- timent classifiers as features in addition to lexical and syntactic features, and per- formed in the top ten of other constrained systems on both domains. 1 Introduction As stated in the call for participation for this Se- meval task, sentiment analysis focusing on overall polarity of a document, sentence, or similar con- text has been well studied in recent years (Liu, 2010; Pang and Lee, 2008; Tsytsarau and Pal- panas, 2012). However, there is less prior work examining finer levels of granularity associated with individual entities and their characteristics or attributes, which the organizers for this task call aspects. The aspect based sentiment analysis This work is licenced under a Creative Commons At- tribution 4.0 International License. Page numbers and pro- ceedings footer are added by the organizers. License de- tails: http://creativecommons.org/licenses/ by/4.0/ task (ABSA) has the goal of identifying aspects of stated or implied target entities and the senti- ment expressed towards each aspect. This prob- lem has not been deeply studied in prior literature due to the lack, until now, of a large gold standard dataset. This Semeval task has provided two such datasets, in the domains of laptops and restaurants. A full description of the task and data is presented with this volume (Pontiki et al., 2014). In this paper, we discuss our approach to the first two subtasks of the Semeval ABSA Task, those of aspect term extraction and aspect term polarity. In aspect term extraction the domain (restaurants or laptops) is known and the goal is to identify terms in a sentence that are features commonly associated with that domain, such as service and staff in the case of restaurants or size and speed in the case of laptops. In the polarity subtask, the aspect terms for a given sentence are already identified and the sentiment polarity (pos- itive, negative, conflict, or neutral) must be as- signed. We approach both subtasks using supervised machine learning with background knowledge of sentiment lexicons and syntax included in our fea- ture set. Our goal …

Content truncated to 500 words