Gupta, Aditi
(2023)
Similarity-Based Weights For
Cross-Domain Sentiment Classification Of Product Reviews.
PhD thesis, Perpustakaan Hamzah Sendut.
Abstract
The unavailability of labelled data for a particular domain poses a challenge for training a classifier for sentiment detection in product reviews. Cross-domain sentiment analysis offers a solution to train models using labelled data from source domains and applying it to the target domain. However, the classifier performance usually suffers significantly when the source and target domains’ feature distribution and sentiment expressions differ. Also, when using multiple source domains, not all source domains are equally beneficial as some are more relevant to a particular target domain. This thesis addresses these issues by developing cross-domain deep learning classifiers and investigating the impact of multiple source domains on sentiment classifier training. Furthermore, the effect of each source domain on the training of the cross-domain sentiment classifier and selecting helpful source domains is examined. The study developed a novel method of assigning weights, to each source domain according to its importance to the target domain. A three-phase methodology is implemented, with Phase 1 focusing on creating the deep learning architecture using CNN with optimal hyperparameters for cross-domain classification tasks followed by extensive experiments to find the relevance between various source domains to the target domain.
Actions (login required)
 |
View Item |
