Prototypical Opinion Holders: What We can Learn from Experts and Analysts

Proceedings of Recent Advances in Natural Language Processing, pages 282–288, Hissar, Bulgaria, 12-14 September 2011. Prototypical Opinion Holders: What We can Learn from Experts and Analysts Michael Wiegand and Dietrich Klakow Spoken Language Systems Saarland University D-66123 Saarbr¨ucken, Germany {Michael.Wiegand|Dietrich.Klakow}@lsv.uni-saarland.de Abstract In order to automatically extract opinion holders, we propose to harness the con- texts of prototypical opinion holders, i.e. common nouns, such as experts or an- alysts, that describe particular groups of people whose profession or occupation is to form and express opinions towards spe- cific items. We assess their effectiveness in supervised learning where these con- texts are regarded as labeled training data and in rule-based classification which uses predicates that frequently co-occur with mentions of the prototypical opinion hold- ers. Finally, we also examine in how far knowledge gained from these contexts can compensate the lack of large amounts of labeled training data in supervised learn- ing by considering various amounts of ac- tually labeled training sets. 1 Introduction Building an opinion holder (OH) extraction sys- tem on the basis of supervised classifiers requires large amounts of labeled training data which are expensive to obtain. Therefore, alternative meth- ods requiring less human effort are required. Such methods would be particularly valuable for lan- guages other than English as for most other lan- guages sentiment resources are fairly sparse. In this paper, we propose to leverage contex- tual information from prototypical opinion holders (protoOHs), such as experts or analysts. We define prototypical opinion holders as common nouns de- noting particular groups of people whose profes- sion or occupation is to form and express opinions towards specific items. Mentions of these nouns are disproportionately often OHs: 1. Experts agree it generally is a good idea to follow the manufacturers’ age recommendations. 2. Shares of Lotus Development Corp. dropped sharply after analysts expressed concern about their business. Since protoOHs are common nouns they should occur sufficiently often in a large text corpus in order to gain knowledge for OH extraction. We examine different ways of harnessing mentions of protoOHs for OH extraction. We compare their usage as labeled training data for supervised learn- ing with a rule-based classifier that relies on a lex- icon of predictive predicates that have been ex- tracted from the contexts of protoOHs. Moreover, we investigate in how far the knowledge gained from these contexts can compensate the lack of large amounts of actually labeled training data in supervised classification by considering various amounts of labeled training sets. 2 Related Work There has been much research on supervised learning for OH extraction. Choi et al. (2005) explore OH extraction using CRFs with several manually defined linguistic features and automat- ically learnt surface patterns. The linguistic fea- tures focus on named-entity information and syn- tactic relations to opinion words. Kim and Hovy (2006) and Bethard et al. (2004) examine the use- fulness of semantic roles provided by FrameNet1 for both OH and opinion target extraction. More recently, Wiegand and Klakow (2010) explored convolution kernels for OH extraction …

Content truncated to 500 words