Structural Alignment for Comparison Detection

Proceedings of Recent Advances in Natural Language Processing, pages 275–281, Hissar, Bulgaria, Sep 7–9 2015. Structural Alignment for Comparison Detection Wiltrud Kessler and Jonas Kuhn Institute for Natural Language Processing University of Stuttgart wiltrud.kessler@ims.uni-stuttgart.de Abstract There tends to be a substantial propor- tion of reviews that include explicit textual comparisons between the reviewed item and another product. To the extent that such comparisons can be captured reliably by automatic means, they can provide an extremely helpful input to support a pro- cess of choice. As the small amount of available training data limits the develop- ment of robust systems to automatically detect comparisons, this paper investigates how to use semi-supervised strategies to expand a small set of labeled sentences. Specifically, we use structural alignment, a method that starts out from a seed set of manually annotated data and finds simi- lar unlabeled sentences to which the labels can be projected. We present several adap- tations of the method to our task of com- parison detection and show that adding the found expansion sentences slightly im- proves over a non-expanded baseline in low-resource settings, i.e., when a very small amount of training data is available. 1 Introduction Sentiment analysis is an NLP task that has re- ceived considerable attention in recent years. If we consider the actual situations in which people are interested in aggregated subjective assessments of some product (or location, service etc.) by other users, a typical scenario is that they are in the pro- cess of making some choice – such as a purchase decision among a set of candidate products. It is clear that for this decision a plain polarity scor- ing for entire review texts is of limited use and we need a more detailed analysis. In this work, we focus on what is presumably the most useful kind of expression when it comes to supporting a pro- cess of choice: there tends to be a substantial pro- portion of reviews (about 10% of sentences) that include explicit textual comparisons, e.g., “X is better than Y”. To the extent that such subjective comparisons can be captured reliably by automatic means, they can provide an extremely helpful ba- sis for coming up with a decision. The analysis of comparisons has the disad- vantage that data for supervised training can no longer be derived from star ratings. Existing man- ually annotated sentiment analysis data sets in- clude some proportion of comparisons, however, for a reliable supervised training, a larger data set is required. Moreover, vocabulary differences across product categories make it advisable to use domain-specific training data. If enough (human and/or financial) resources are available, the most effective approach is of course to invest in quality-controlled manual an- notation of a relatively large amount of training data. However, since the higher-level semantic structure of comparisons as they appear in reviews is clear-cut, the problem setting could respond fa- vorably to weakly supervised training strategies that start out from a seed set of manually annotated data. The experiments we …

Content truncated to 500 words