Bootstrapping Polar-Opposite Emotion Dimensions from Online Reviews

Bootstrapping Polar-Opposite Emotion Dimensions from Online Reviews Luwen Huangfu and Mihai Surdeanu University of Arizona Tucson, AZ 85719, USA {huangfuluwen, msurdeanu}@email.arizona.edu Abstract We propose a novel bootstrapping approach for the acquisition of lexicons from unannotated, informal online texts (in our case, Yelp reviews) for polar-opposite emotion dimension values from the Ortony/Clore/Collins model of emotions (e.g., desirable/undesirable). Our approach mitigates the intrinsic problem of limited supervision in bootstrapping with an effective strategy that softly labels unlabeled terms, which are then used to better estimate the quality of extraction patterns. Further, we propose multiple solutions to control for semantic drift by taking advantage of the polarity of the categories to be learned (e.g., praiseworthy vs. blameworthy). Experimental results demonstrate that our algorithm achieves considerably better performance than several baselines. Keywords: bootstrapping, semantic drift, limited supervision 1. Introduction Bootstrapping is a lightly-supervised learning approach in which supervision comes in the form of a small number of initial examples (or seeds). While bootstrapping is an attractive choice for NLP, the limited supervision involved also yields important drawbacks. First, the training of boot- strapping models often “drifts” semantically from the orig- inal task into different tasks (e.g., from learning women names into learning flower names). Second, the lack of la- beled data (i.e., only a small set of seed examples is anno- tated) impedes the capacity of the model to correctly assess the quality of the generated model during training. In this paper we propose solutions for the above issues in the context of learning lexicons for the emotion dimen- sions (e.g., Desirability, Praise-/Blame-worthiness) neces- sary to assemble the Ortony/Clore/Collins (OCC) cognitive model of emotions (1990). According to the OCC frame- work, 22 different emotion types are generated from several emotion dimensions. In this work, we focus on the dimen- sions of Desirability, Praise-/Blame-worthiness and Like- lihood because they are central emotion dimensions that contain both positive values (e.g., desirable, praiseworthy, certain) and negative values (e.g., undesirable, blamewor- thy, likely). For example, the emotion type JOY combines the dimensions Desirability with value desirable and Like- lihood with value certain. The contributions of this work are: 1. We propose multiple lightly-supervised solutions for the acquisition of emotion dimensions that control for semantic drift by taking advantage of the polarity of the classes to be learned (i.e., positive/negative ap- praisals). 2. We introduce an effective strategy to softly label unla- beled terms, i.e., unlabeled terms are assigned a value that indicates how close they are to a given category, and use these soft labels to better estimate the quality of extraction patterns in the above bootstrapping ap- proaches. 3. We show that multiple resources (WordNet, word em- beddings that project words in a continuous vector space that capture distributional similarity, and edit distance similarity) all help for the above two contri- butions, and are complementary to each other. 4. We empirically demonstrate that our approach outper- forms several strong baselines for the acquisition of emotion dimensions lexicons from informal texts such as product reviews on the web. …

Content truncated to 500 words