A Weakly Supervised Model for Sentence-Level Semantic Orientation Analysis with Multiple Experts

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 149–159, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics A Weakly Supervised Model for Sentence-Level Semantic Orientation Analysis with Multiple Experts Lizhen Qu and Rainer Gemulla and Gerhard Weikum Max Planck Institute for Informatics Saarbr¨ucken, Germany {lqu,rgemulla,weikum}@mpi-inf.mpg.de Abstract We propose the weakly supervised Multi- Experts Model (MEM) for analyzing the se- mantic orientation of opinions expressed in natural language reviews. In contrast to most prior work, MEM predicts both opinion po- larity and opinion strength at the level of in- dividual sentences; such fine-grained analysis helps to understand better why users like or dislike the entity under review. A key chal- lenge in this setting is that it is hard to ob- tain sentence-level training data for both po- larity and strength. For this reason, MEM is weakly supervised: It starts with potentially noisy indicators obtained from coarse-grained training data (i.e., document-level ratings), a small set of diverse base predictors, and, if available, small amounts of fine-grained train- ing data. We integrate these noisy indicators into a unified probabilistic framework using ideas from ensemble learning and graph-based semi-supervised learning. Our experiments in- dicate that MEM outperforms state-of-the-art methods by a significant margin. 1 Introduction Opinion mining is concerned with analyzing opin- ions expressed in natural language text. For example, many internet websites allow their users to provide both natural language reviews and numerical ratings to items of interest (such as products or movies). In this context, opinion mining aims to uncover the relationship between users and (features of) items. Preferences of users to items can be well understood by coarse-grained methods of opinion mining, which focus on analyzing the semantic orientation of doc- uments as a whole. To understand why users like or dislike certain items, however, we need to perform more fine-grained analysis of the review text itself. In this paper, we focus on sentence-level analy- sis of semantic orientation (SO) in online reviews. The SO consists of polarity (positive, negative, or other1) and strength (degree to which a sentence is positive or negative). Both quantities can be ana- lyzed jointly by mapping them to numerical ratings: Large negative/positive ratings indicate a strong neg- ative/positive orientation. A key challenge in fine- grained rating prediction is that fine-grained train- ing data for both polarity and strength is hard to obtain. We thus focus on a weakly supervised set- ting in which only coarse-level training data (such as document ratings and subjectivity lexicons) and, optionally, a small amount of fine-grained training data (such as sentence polarities) is available. A number of lexicon-based approaches for phrase- level rating prediction has been proposed in the liter- ature (Taboada et al., 2011; Qu et al., 2010). These methods utilize a subjectivity lexicon of words along with information about their semantic orientation; they focus on phrases that contain words from the lexicon. A key advantage of sentence-level methods is that they are able …

Content truncated to 500 words