Speaker: Areej Alhothali, PhD Candidate
Natural language texts are often meant to express or impact individuals emotions. Sentiment analysis researchers are increasingly interested in investigating natural language processing techniques as well as emotions theories to classify sentiments expressed in natural language text. Most sentiment analysis research effort focuses on classifying highly opinionated documents from the writer's perspectives and uses either count-based word representations that ignore sentence structure or a small set of lexicon resources that do not cover the wide range of words used on the Internet.
In this thesis, we propose a cross-disciplinary approach that incorporates Affect Control Theory (ACT) into a structured model to determine three-dimensional sentiment polarity of event-based articles from the readers and interactants perspectives. ACT is a socio-mathematical theory that models humans' sentiment towards social events in a three-dimensional space and estimates the situational post-event sentiment by considering the interaction between these sentiments.
We develop an algorithm that performs a fine-grained events extraction from English sentences using a combination of semantic and syntactic parsing techniques. We also augment an ACT three-dimensional lexicon in a semi-supervised fashion using a graph-based label propagation method built upon semantic and distributed word representations. Using the extracted events, the augmented lexicon, and the ACT mathematical equations, we propose an entity-based sentiment analysis approach that models the readers triggered emotions from event-based sentences toward the events and their associated entities. We also propose an ACT-based model to predict the temporal progression of the interactants emotions and their optimal behaviour (acts) over a sequence of interactions. We evaluate the first approach on a manually annotated news headline corpus and the second on three different types of corpora: fairy tales, news articles, and a hand-crafted corpus.
The results demonstrate that, despite the challenging structure of the sentences, there was a reasonable agreement between the estimated emotions and behaviours and their corresponding ground truths.