Document graph representation learning

• Main Author: ZHANG, Ce
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2023
• Subjects: Topic Modeling; Text Mining; Graph Representation Learning; Graph Neural Networks; Graphics and Human Computer Interfaces; OS and Networks
• Online Access: https://ink.library.smu.edu.sg/etd_coll/496; https://ink.library.smu.edu.sg/context/etd_coll/article/1494/viewcontent/GPIS_AY2018_PhD_Ce_Zhang.pdf
• Institution: Singapore Management University

Much of the data on the Web can be represented in a graph structure, ranging from social and biological to academic and Web page graphs, etc. Graph analysis recently attracts escalating research attention due to its importance and wide applicability. Diverse problems could be formulated as graph tasks, such as text classification and information retrieval. As the primary information is the inherent structure of the graph itself, one promising direction known as the graph representation learning problem is to learn the representation of each node, which could in turn fuel tasks such as node classification, node clustering, and link prediction. As a specific graph data, documents are usually connected in a graph structure. For example, Google Web pages hyperlink to other related pages, academic papers cite other papers, Facebook user profiles are connected as a social network, news articles with similar tags are linked together, etc. We call such data document graph or document network. To better make sense of the meaning within these text documents, researchers develop neural topic models. By modeling both textual content within documents and connectivity across documents, we can discover more interpretable topics to understand the corpus and better fulfill real-world applications, such as Web page searching, news article classification, academic paper indexing, and friend recommendation based on user profiles, etc. However, traditional topic models explore the content only, ignoring the connectivity. In this dissertation, we aim to develop models for document graph representation learning. First, we investigate the extension of Auto-Encoders, a family of shallow topic models. Intuitively, connected documents tend to share similar latent topics. Thus, we allow Auto-Encoder to extract topics of the input document and reconstruct its adjacent neighbors. This allows documents in a network to collaboratively learn from one another, such that close neighbors would have similar representations in the topic space. Extensive experiments verify the effectiveness of our proposed model against both graphical and neural baselines. Second, we focus on dynamic modeling of document networks. In many real-world scenarios, documents are published in a sequence and are associated with timestamps. For example, academic papers published over the years exhibit the development of research topics. To incorporate such temporal information, we introduce a neural topic model aimed at learning unified topic distributions that incorporate both document dynamics and network structure. Third, we discover that documents are usually associated with authors. For example, news reports have journalists specializing in writing certain type of events, academic papers have authors with expertise in certain research topics, etc. Modeling authorship information could benefit topic modeling, since documents by the same authors tend to reveal similar semantics. This observation also holds for documents published on the same venues. We propose a Variational Graph Author Topic Model for documents to integrate both topic modeling and authorship and venue modeling into a unified framework. Fourth, most previous topic models treat documents of different lengths uniformly, assuming that each document is sufficiently informative. However, shorter documents may have only a few word co-occurrences, resulting in inferior topic quality. Some other previous works assume that all documents are short, and leverage external auxiliary data, e.g., pretrained word embeddings and document connectivity. Orthogonal to existing works, we remedy this problem within the corpus itself by meta-learning and proposing a Meta-Complement Topic Model, which improves topic quality of short texts by transferring the semantic knowledge learned on long documents to complement semantically limited short texts. Fifth, we explore the modeling of short texts on the graph. Text embedding models usually rely on word co-occurrences within the documents to learn effective representations. However, short texts with only a few words may influence the learning process. To accurately discover the main topics of these short documents, we propose a new statistical concept, i.e., optimal transport barycenter, to incorporate external knowledge, such as pre-trained word embedding on a large corpus, to improve topic modeling. The proposed model shows state-of-the-art performance.