Domain-Adaptive Sentiment Analysis Across Online Social Networks
Aspect-based sentiment analysis is an important task in natural language processing and has a wide range of applications in fields such as e-commerce, marketing, and customer service. The goal of this task is to identify aspect and opinion terms and classify the sentiment expressed towards a particular aspect in a given text. Despite its significance, aspect-based sentiment analysis remains a challenging task due to limitations in existing models. These limitations include an inadequate consideration of crucial implicit linguistic features for aspect term extraction, declining performance on unstructured and small datasets for aspect and relation extraction, a complex and varied model landscape for different sub-tasks, and the time-consuming construction of prompts for cross-domain aspect term extraction. In this thesis, these challenges are tackled by employing several innovative deep neural network models.
First, a novel and efficient framework is introduced for extracting aspect terms by combining contextual and linguistic features using the Artificial Bee Colony-based feature selection method. To address the high sparsity and dimensionality of raw data, an improved version of Artificial Bee Colony is employed to determine the most relevant linguistic features. These selected features and context embeddings are then integrated to enhance the accuracy of aspect extraction.
Second, a novel and deep learning-based model is proposed for recognising individuals’ concerns and the associated relationships through the integration of Graph Convolutional Networks, Bi-directional Long Short-Term Memory, and Concern Graphs. The proposed model leverages sequential features from BERT embeddings and regional features of tweets extracted through the Concern Graph module, leading to improved concern detection and high resistance to noise. This approach overcomes the limitations of limited manually labelled data.
Third, a novel integrated framework is designed to tackle all defined sub-tasks of aspect-based sentiment analysis. The framework consists of a multi-layer semantic model based on graph convolutional networks, which is designed to capture the semantic connections between aspects and opinion terms. Additionally, a multi-layer syntax model is proposed to learn explicit dependency relations at different levels. To support the sub-tasks, the semantic features learned by the semantic model are passed to the syntax model, providing enhanced semantic guidance and allowing the syntax model to learn more comprehensive syntactic representations. The framework incorporates two attention mechanisms, one for modelling dependency relations and types and another for encoding part-of-speech tags to detect aspect and opinion term boundaries. This differs from conventional syntactic models, making the proposed framework unique.
Finally, a novel soft prompt-based joint learning method is proposed for aspect term extraction across domains. The method utilises external linguistic features to learn domain-invariant representations between source and target domains through multiple objectives, effectively bridging the gap between domains with varying distributions of aspect terms. Furthermore, the method incorporates a set of transferable soft prompts, consisting of multiple learnable vectors, to enhance the detection of aspect terms in the target domain. The proposed soft prompt-based joint learning method represents a novel approach to cross-domain aspect term extraction.