Unsupervised Topic-Specific Domain Dependency Graphs for Aspect Identification in Sentiment Analysis

We propose to model a collection of documents by means of topic-specific domain dependency graphs (DDGs). We use LDA topic modeling to detect topics underlying a mixed-domain dataset and select topically pure documents from the collection. We aggregate counts of words and their dependency relations per topic, weigh them with Tf-Idf and produce a DDG by selecting the highest-ranked words and their dependency relations. We demonstrate an implementation of the approach on the task of identifying product aspects for aspect-oriented sentiment analysis. A large corpus of Amazon reviews is used to identify product aspects by applying syntactic filtering to the DDG. Evaluation on a small set of cameras reviews demonstrate a good precision of our method. To our knowledge, this is the first method that finds product-class specific aspects in mutli-domain collections in an unsupervised fashion.