Multiple Aspect Ranking Using the Good Grief Algorithm

Proceedings of NAACL HLT 2007, pages 300–307, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Multiple Aspect Ranking using the Good Grief Algorithm Benjamin Snyder and Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {bsnyder,regina}@csail.mit.edu Abstract We address the problem of analyzing mul- tiple related opinions in a text. For in- stance, in a restaurant review such opin- ions may include food, ambience and ser- vice. We formulate this task as a multiple aspect ranking problem, where the goal is to produce a set of numerical scores, one for each aspect. We present an algorithm that jointly learns ranking models for in- dividual aspects by modeling the depen- dencies between assigned ranks. This al- gorithm guides the prediction of individ- ual rankers by analyzing meta-relations between opinions, such as agreement and contrast. We prove that our agreement- based joint model is more expressive than individual ranking models. Our empirical results further confirm the strength of the model: the algorithm provides significant improvement over both individual rankers and a state-of-the-art joint ranking model. 1 Introduction Previous work on sentiment categorization makes an implicit assumption that a single score can express the polarity of an opinion text (Pang et al., 2002; Turney, 2002; Yu and Hatzivassiloglou, 2003). However, multiple opinions on related matters are often intertwined throughout a text. For example, a restaurant review may express judgment on food quality as well as the service and ambience of the restaurant. Rather than lumping these aspects into a single score, we would like to capture each aspect of the writer’s opinion separately, thereby providing a more fine-grained view of opinions in the review. To this end, we aim to predict a set of numeric ranks that reflects the user’s satisfaction for each as- pect. In the example above, we would assign a nu- meric rank from 1-5 for each of: food quality, ser- vice, and ambience. A straightforward approach to this task would be to rank1 the text independently for each aspect, us- ing standard ranking techniques such as regression or classification. However, this approach fails to ex- ploit meaningful dependencies between users’ judg- ments across different aspects. Knowledge of these dependencies can be crucial in predicting accurate ranks, as a user’s opinions on one aspect can influ- ence his or her opinions on others. The algorithm presented in this paper models the dependencies between different labels via the agreement relation. The agreement relation captures whether the user equally likes all aspects of the item or whether he or she expresses different degrees of satisfaction. Since this relation can often be deter- mined automatically for a given text (Marcu and Echihabi, 2002), we can readily use it to improve rank prediction. The Good Grief model consists of a ranking model for each aspect as well as an agreement model which predicts whether or not all rank aspects are 1In this paper, ranking refers to the task of assigning an inte- ger from 1 to k to each instance. …

Content truncated to 500 words