Graph Fusion Network for Text Classification

Dai, Yong; Shou, Linjun; Gong, Ming; Xia, Xiaolin; Kang, Zhao; Xu, Zenglin; Jiang, Daxin

doi:10.1016/j.knosys.2021.107659

Cited by 39 publications

(15 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Recently, graph attention mechanism has been applied in text classification tasks (Mei et al, 2021;Yang et al, 2021a). Others focus on using both local and global information (Jin et al, 2021), multi-modality with text and image information (Yang et al, 2021b), enhancing TextGCN with other models (Ragesh et al, 2021) and combining with external knowledge (Dai et al, 2022).…”

Section: Related Workmentioning

confidence: 99%

ME-GCN: Multi-dimensional Edge-Embedded Graph Convolutional Networks for Semi-supervised Text Classification

Wang¹,

Han²,

Long³

et al. 2022

Preprint

View full text Add to dashboard Cite

Compared to sequential learning models, graph-based neural networks exhibit excellent ability in capturing global information and have been used for semisupervised learning tasks. Most Graph Convolutional Networks are designed with the single-dimensional edge feature and failed to utilise the rich edge information about graphs. This paper introduces the ME-GCN (Multi-dimensional Edgeenhanced Graph Convolutional Networks) for semi-supervised text classification. A text graph for an entire corpus is firstly constructed to describe the undirected and multi-dimensional relationship of word-to-word, document-document, and word-to-document. The graph is initialised with corpus-trained multi-dimensional word and document node representation, and the relations are represented according to the distance of those words/documents nodes. Then, the generated graph is trained with ME-GCN, which considers the edge features as multi-stream signals, and each stream performs a separate graph convolutional operation. Our ME-GCN can integrate a rich source of graph edge information of the entire text corpus. The results have demonstrated that our proposed model has significantly outperformed the state-of-the-art methods across eight benchmark datasets. The code is available on: https://github.com/usydnlp/ME GCN

show abstract

Section: Related Workmentioning

confidence: 99%

ME-GCN: Multi-dimensional Edge-Embedded Graph Convolutional Networks for Semi-supervised Text Classification

Wang¹,

Han²,

Long³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…From an architecture perspective, conventional sequence-based or convolutional neural networks that are often utilized for text classification are limited by their nature to prioritize sequentiality and locality [85,86]. While these deep learning models capture semantic and syntactic information in the Euclidean space and in local sequences well, they do not account for global word co-occurrences in a corpus that carries non-consecutive and long-distance semantics [39,87].…”

Section: Convolutional Neural Networkmentioning

confidence: 99%

“…This section reviews the graph convolutional network (GCN). In recent years, features for text classification have been generated from non-Euclidean domains and are represented in the form of graphs [87]. These techniques preserve diverse global structural information and capture multi-dimension relational information as meaningful features [88].…”

Section: Graph Networkmentioning

confidence: 99%

“…As an extension to sequential-based graphs [89], a TensorGCN triple-graph model has been developed to describe syntactic, semantic, and sequential information among word nodes [40]-in line with SynGCN and SemGCN in terms of the need to incorporate embeddings beyond sequentiality [90]. More recently, the graph fusion network (GFN) addresses the limitation of transductive methods [91] in adapting to new documents by discarding document nodes and constructing four homogeneous graphs [87]. Since the majority of graph models for text classification only consider word nodes [39,40], direct application of these methods is not suitable for domain-agnostic text classification-they lack representations associated with additional textual features such as phrases, regexes, or topics.…”

Section: Graph Networkmentioning

confidence: 99%

“…Here, p intra and p inter denote the application of each learning method and H (l) intra denotes the graph feature tensor after performing intra-graph propagation by applying (2.19) on A and H. Unlike (2.19), the trainable weight matrix W (l,g) intra is designed to be graphspecific with g denoting the gth graph. To achieve inter-graph propagation, on the other hand, a series of virtual graphs are constructed by duplicating the same set of nodes and In contrast to TensorGCN, GFN [87] consists of four homogeneous word graphs with edge links comprising word embedding cosine similarities and Euclidean distances, PMI, and co-occurrence statistics. To generate document embeddings without document nodes, GFN implements the late fusion paradigm (i.e., the logit-level fusion of word embeddings) [87].…”

Section: Graph Networkmentioning

confidence: 99%

See 2 more Smart Citations

Domain-agnostic document and question classification using natural language processing techniques

Supraja¹

View full text Add to dashboard Cite

I would like to dedicate this thesis to two special angels: God and my dear parents, without whom I could not have completed this challenging, yet fruitful PhD journey. I have no words to express my gratitude for every single thing my parents have done for my education and for my whole life. I cannot do anything to repay their love, affection, care, and concern for me.I would like to extend my sincere thanks to my husband, all my family members and friends who have supported me in one way or another. First and foremost, I would like to sincerely express my gratitude to my advisor A/Prof Andy W. H. Khong for his continuous support and guidance throughout my Ph.D. journey. His constant motivation and encouragement has transformed me tremendously.His advice has always been a great tonic to boost my morale and confidence level.

show abstract