Use this URL to cite or link to this record in EThOS:
Title: Adaptive sentiment analysis
Author: Mudinas, Andrius
ISNI:       0000 0004 7967 9123
Awarding Body: Birkbeck, University of London
Current Institution: Birkbeck (University of London)
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Domain dependency is one of the most challenging problems in the field of sentiment analysis. Although most sentiment analysis methods have decent performance if they are targeted at a specific domain and writing style, they do not usually work well with texts that are originated outside of their domain boundaries. Often there is a need to perform sentiment analysis in a domain where no labelled document is available. To address this scenario, researchers have proposed many domain adaptation or unsupervised sentiment analysis methods. However, there is still much room for improvement, as those methods typically cannot match conventional supervised sentiment analysis methods. In this thesis, we propose a novel aspect-level sentiment analysis method that seamlessly integrates lexicon- and learning-based methods. While its performance is comparable to existing approaches, it is less sensitive to domain boundaries and can be applied to cross-domain sentiment analysis when the target domain is similar to the source domain. It also offers more structured and readable results by detecting individual topic aspects and determining their sentiment strengths. Furthermore, we investigate a novel approach to automatically constructing domain-specific sentiment lexicons based on distributed word representations (aka word embeddings). The induced lexicon has quality on a par with a handcrafted one and could be used directly in a lexiconbased algorithm for sentiment analysis, but we find that a two-stage bootstrapping strategy could further boost the sentiment classification performance. Compared to existing methods, such an end-to-end nearly-unsupervised approach to domain-specific sentiment analysis works out of the box for any target domain, requires no handcrafted lexicon or labelled corpus, and achieves sentiment classification accuracy comparable to that of fully supervised approaches. Overall, the contribution of this Ph.D. work to the research field of sentiment analysis is twofold. First, we develop a new sentiment analysis system which can - in a nearlyunsupervised manner-adapt to the domain at hand and perform sentiment analysis with minimal loss of performance. Second, we showcase this system in several areas (including finance, politics, and e-business), and investigate particularly the temporal dynamics of sentiment in such contexts.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available