Automatic Topic Model Adaptation for Sentiment Analysis in Structured Domains

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 75–83, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Automatic Topic Model Adaptation for Sentiment Analysis in Structured Domains Geoffrey Levine and Gerald DeJong University of Illinois at Urbana-Champaign Urbana, IL levine@illinois.edu and mrebl@illinois.edu Abstract We present a novel topic modeling ap- proach to sentiment analysis for docu- ments organized into hierarchical cate- gories. In our approach, positive, nega- tive, and subject matter topics are learned and used to infer document labels. Markov chain Monte Carlo model pro-A cedure adapts the number and structure of topics based on a minimum descrip- tion length objective function. We ap- ply our approach to Yelp.com business re- views and Amazon.com book reviews and demonstrate that 1) the model adaptation procedure selects a high quality model from the space of alternatives, and 2) the resulting model performs well relative to state of the art regression and topic model- ing approaches. 1 Introduction Selecting an appropriate model is an important part of any machine learning endeavor. The model must be chosen in a manner so as to balance two objectives: 1) Be sufficiently rich to capture the relevant patterns in the data, and 2) Be simple enough to avoid spurious patterns in the training data (overfitting). In natural language processing tasks, there are often many modeling choices to be made regarding what feature granularities and interactions to consider. It is important to make these decisions in a manner such that the resulting model strikes a balance between these two some- what contradictory objectives. In order to appropriately make these choices, we must consider not only the task involved but also the training data available. With copious data we can reliably calibrate complex models, but with limited data complex models risk overfitting. Many general model selection techniques exist in which each candidate model is fit to the training data and scored with respect to a particular cri- terion. While these approaches allow us to com- pare a small number of models in order to select the most appropriate, they require calibrating each model’s parameters to the training data. However, when there are many modeling choices to be made and thus a large space of alternative models, fitting all of them to the training data is computationally prohibitive. In this paper, we present a novel topic mod- eling approach for structured sentiment analysis domains and an automatic model adaptation ap- proach that takes advantage of categorical meta- data. This model adaptation approach resolves the structure of the metadata with the significant pat- terns in the training data to determine the number and range of latent topics. We demonstrate our approach on Yelp.com business reviews as well as Amazon.com book re- views. We show that our model adaptation ap- proach selects an appropriate model given a par- ticular amount of training data, and the resulting model is high quality relative to alternative regres- sion and topic modeling approaches. 2 Background Sentiment analysis, in which …

Content truncated to 500 words