Extracting Opinion Targets in a Single and Cross-Domain Setting with Conditional Random Fields

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1035–1045, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Extracting Opinion Targets in a Single- and Cross-Domain Setting with Conditional Random Fields Niklas Jakob Technische Universit¨at Darmstadt Hochschulstraße 10 64289 Darmstadt, Germany Iryna Gurevych Technische Universit¨at Darmstadt Hochschulstraße 10 64289 Darmstadt, Germany http://www.ukp.tu-darmstadt.de/people Abstract In this paper, we focus on the opinion tar- get extraction as part of the opinion min- ing task. We model the problem as an in- formation extraction task, which we address based on Conditional Random Fields (CRF). As a baseline we employ the supervised al- gorithm by Zhuang et al. (2006), which rep- resents the state-of-the-art on the employed data. We evaluate the algorithms comprehen- sively on datasets from four different domains annotated with individual opinion target in- stances on a sentence level. Furthermore, we investigate the performance of our CRF-based approach and the baseline in a single- and cross-domain opinion target extraction setting. Our CRF-based approach improves the perfor- mance by 0.077, 0.126, 0.071 and 0.178 re- garding F-Measure in the single-domain ex- traction in the four domains. In the cross- domain setting our approach improves the per- formance by 0.409, 0.242, 0.294 and 0.343 re- garding F-Measure over the baseline. 1 Introduction The automatic extraction and analysis of opinions has been approached on several levels of granular- ity throughout the last years. As opinion mining is typically an enabling technology for another task, this overlaying system defines requirements regard- ing the level of granularity. Some tasks only require an analysis of the opinions on a document or sen- tence level, while others require an extraction and analysis on a term or phrase level. Amongst the tasks which require the finest level of granularity are: a) Opinion question answering - i.e. with ques- tions regarding an entity as in “What do the people like / dislike about X?”. b) Recommender systems - i.e. if the system shall only recommend entities which have received good reviews regarding a cer- tain aspect. c) Opinion summarization - i.e. if one wants to create an overview of all positive / negative opinions regarding aspect Y of entity X and cluster them accordingly. All of these tasks have in com- mon that in order to fulfill them, the opinion min- ing system must be capable of identifying what the opinions in the individual sentences are about, hence extract the opinion targets. Our goal in this work is to extract opinion tar- gets from user-generated discourse, a discourse type which is quite frequently encountered today, due to the explosive growth of Web 2.0 community web- sites. Typical sentences which we encounter in this discourse type are shown in the following examples. The opinion targets which we aim to extract are un- derlined in the sentences, the corresponding opinion expressions are shown in italics. (1) While none of the features are earth-shattering, eCircles does provide a great place to keep in touch. (2) Hyundai’s more-than-modest …

Content truncated to 500 words