Core Concepts
Relational Graph Convolutional Networks (RGCNs) can effectively capture complex relationships between entities in text data to enhance sentiment analysis performance.
Abstract
The paper proposes leveraging Relational Graph Convolutional Networks (RGCNs) for sentiment analysis tasks. The key highlights are:
Construction of a heterogeneous graph from the text corpus, with three types of edges: word-word (based on co-occurrence), word-document (based on TF-IDF), and document-document (based on Jaccard similarity).
Initialization of node embeddings using pre-trained language models like BERT and RoBERTa, which capture rich semantic information.
Application of the RGCN framework to the constructed graph, which can effectively model diverse relationships between words, documents, and other entities.
Experiments on two diverse datasets - English Amazon reviews and Persian Digikala reviews - demonstrate the superior performance of the RGCN-based approach compared to traditional methods and standard GCNs.
The RGCN model outperforms BERT and RoBERTa baselines, as well as GCN-based approaches, by effectively capturing the relational information in the text data.
The results highlight the advantages of RGCNs in handling complex relationships and diverse interactions, leading to improved sentiment analysis performance.
Stats
The Amazon dataset has 151,232 sentences, with 5,774 sentences in class 1, 7,907 in class 2, 17,490 in class 3, 32,777 in class 4, and 87,284 in class 5.
The Digikala dataset has 63,586 sentences, with 16,098 in the "Not Recommended" class, 10,528 in the "No Opinion" class, and 36,960 in the "Recommended" class.
Quotes
"Graph Neural Networks (GNNs) have emerged as a powerful paradigm for analyzing structured data, offering unique advantages in capturing relationships and dependencies between data points represented as nodes in a graph."
"Relational Graph Convolutional Networks (RGCNs) address this by using different types of edges to capture different relationships. However, this expressiveness comes at a computational cost."