With the growing popularity and availability of opinion-rich resources such as social media platforms and networks, new opportunities arise as people can now share their opinions and also seek or understand the opinion of others about a specific topic or event. This growth has fuelled interest in opinion mining which seeks to understand opinions, attitudes, judgements and evaluations with respect to an entity or its aspects. The proliferation of reviews, ratings and online expressions have turned into a valuable asset to businesses seeking to manage their reputation, market their products, or identify new opportunities through opinion analysis. On the side of consumers, opinion mining serves as an information source that can support decision making. In this research, we focus on some fundamental challenges in opinion mining and make three contributions. First, we develop a curated corpus for training and evaluating opinion mining models. This corpus annotates sentiment and topic information at both sentence and review levels. It also captures the sentiment and topic time-variance information of the reviews. We demonstrate through experiments that this dataset supports opinion mining tasks such as contrastive opinion mining, and joint sentence and document level sentiment and topic analysis. As the corpus has a time-variance characteristic, it could also support studies in sentiment/topic dynamic analysis. Second, we propose a model for mining contrastive opinion from textual data (contraLDA). Unlike existing models that require input data to be separated into different collections beforehand, contraLDA models contrastive opinion from both single and multiple text collections. The model can also be flexibly trained in weakly-supervised and fully-supervised settings. In addition, the contraLDA model not only mines contrastive opinion but also quantifies the strength of opinion contrastiveness towards the topic of interest. The contraLDA model extracts relevant sentences related to the topics, making sentiment-bearing topics more interpretable. Third, we present an aspect extraction method which integrates a Natural Language Processing (NLP) algorithm and word embedding model to identify implicit and explicit aspect expressions from texts. Unlike existing systems, the proposed approach also maps aspect expressions to their corresponding aspect categories. This process allows easy identification of sentences about different aspects of a product. We demonstrate that this unsupervised approach is comparable to state-of-the-art models.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:752674 |
Date | January 2018 |
Creators | Ibeke, Emmanuel Ebuka |
Contributors | Lin, Chenghua ; Wyner, Adam |
Publisher | University of Aberdeen |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://digitool.abdn.ac.uk:80/webclient/DeliveryManager?pid=237773 |
Page generated in 0.001 seconds