Graph Convolutional Networks (GCNs) have shown strong performance in learning text representations for various tasks such as text classification, due to its expressive power in modeling graph structure data (e.g., a literature citation network). Most existing GCNs are limited to deal with documents included in a pre-defined graph, i.e., it cannot be generalized to out-of-graph documents. To address this issue, we propose to transform the document graph into a word graph, to decouple data samples (i.e., documents in training and test sets) and a GCN model by using a document-independent graph. Such word-level GCN could therefore naturally inference out-of-graph documents in an inductive way. The proposed Word-level Graph (WGraph) can not only implicitly learning word presentation with commonly-used word co-occurrences in corpora, but also incorporate extra global semantic dependency derived from inter-document relationships (e.g., literature citations). An inductive Word-grounded Graph Convolutional Network (WGCN) is proposed to learn word and document representations based on WGraph in a supervised manner. Experiments on text classification with and without citation networks evidence that the proposed WGCN model outperforms existing methods in terms of effectiveness and efficiency.

翻译：图卷积网络（GCNs）因其在建模图结构数据（如文献引用网络）方面的强大表达能力，在文本分类等文本表示学习任务中展现出优异性能。然而，现有大多数GCNs仅能处理预定义图中的文档，无法泛化至图外文档。针对这一问题，本文提出将文档图转化为词汇图，通过构建与文档独立的图结构，实现数据样本（即训练集和测试集中的文档）与GCN模型的解耦。这种基于词汇级别的GCN能够以归纳方式自然地推断图外文档。所提出的词汇级图（WGraph）不仅能通过语料库中常用的词共现信息隐式学习词汇表示，还能融入来自文档间关系（如文献引用）的额外全局语义依赖。在此基础上，我们提出一种归纳式基于词汇的图卷积网络（WGCN），以监督方式基于WGraph学习词汇和文档表示。在有/无引用网络的文本分类实验结果表明，所提出的WGCN模型在有效性和效率方面均优于现有方法。