Grammatical structures for word-level sentiment detection

Existing work in fine-grained sentiment anal- ysis focuses on sentences and phrases but ig- nores the contribution of individual words and their grammatical connections. This is because of a lack of both (1) annotated data at the word level and (2) algorithms that can leverage syn- tactic information in a principled way. We ad- dress the first need by annotating articles from the information technology business press via crowdsourcing to provide training and testing data. To address the second need, we propose a suffix-tree data structure to represent syntac- tic relationships between opinion targets and words in a sentence that are opinion-bearing. We show that a factor graph derived from this data structure acquires these relationships with a small number of word-level features. We demonstrate that our supervised model per- forms better than baselines that ignore syntac- tic features and constraints.