Back

Graph Convolutional Networks for Text Classification

Liang YaoChengsheng MaoYuan Luo

article2018AAAI2,128 citations

Introduces Text GCN, a framework that models an entire text corpus as a graph of word co-occurrences and document-word relations to classify documents without external embeddings, outperforming existing methods especially when labeled training data is scarce.

Listen

Text classification is a critical capability across many automated operations, including spam detection, document routing, opinion mining, and clinical record analysis. Standard deep learning models such as convolutional and recurrent neural networks primarily capture local, sequential word order but often fail to model broader, corpus-wide word co-occurrence relationships. This article evaluates whether framing an entire text collection as a single, heterogeneous graph network can outperform conventional deep learning techniques and reduce reliance on pre-trained external word representations.

The article introduces and evaluates Text Graph Convolutional Networks (Text GCN), a method that converts document classification into a node classification task. The approach builds a single unified graph where words and documents serve as interconnected nodes. Document-to-word relationships are weighted by standard statistical term importance, while word-to-word relationships are established via global co-occurrence across text windows. A two-layer graph convolutional network then jointly learns embeddings for both words and documents using only the training labels. The researchers tested this framework across five standard benchmark datasets spanning news articles, medical literature, and short movie reviews, comparing it against multiple established baselines.

The findings show that Text GCN outperforms competing baseline models on four out of five datasets, demonstrating particular strength on longer documents. For example, on the Ohsumed medical dataset, Text GCN achieved 68.36% accuracy, exceeding all baselines including convolutional and recurrent networks. Second, the method demonstrates high resilience when labeled data is scarce: on the R8 dataset, Text GCN achieved 88.30% accuracy using only 1% of the labeled training data, surpassing several baselines trained on the entire dataset. Third, Text GCN successfully generates highly interpretable and discriminative representations for words and documents without requiring pre-trained embeddings or external knowledge bases. Finally, the model proved less effective on single-sentence sentiment data (76.74% on the movie review dataset), where local word order and syntax dominate and sparse text structures limit graph edge formation.

These results demonstrate that organizations can achieve state-of-the-art text categorization performance while substantially reducing the cost and effort of manual data labeling. Because the model operates effectively in low-data regimes without requiring external linguistic datasets, it minimizes deployment friction for specialized, proprietary, or domain-specific vocabularies. However, practitioners should be aware of a fundamental technical trade-off: the model is inherently transductive, meaning the entire dataset graph (including unlabeled test documents) is required during training, making immediate real-time classification of new, unseen documents impractical without retraining or architectural extensions.

Organizations handling large document collections or domain-specific classification tasks should consider piloting graph-based classification, particularly where labeled training data is limited. Before deploying in real-time operational workflows, future development must focus on extending the architecture to inductive settings, incorporating graph attention mechanisms, or evaluating fast sampling frameworks to support on-demand inference on new documents.

Northwestern Universityclass labelsconvolutional neural networkstext classificationText Graph Convolutional Networksgraph convolutional networkstext classification
arXiv: 1809.05679yao8839836/text_gcn
Cover for Graph Convolutional Networks for Text Classification

Abstract

Text classification is an important and classical problem in natural language processing. There have been a number of studies that applied convolutional neural networks (convolution on regular grid, e.g., sequence) to classification. However, only a limited number of studies have explored the more flexible graph convolutional neural networks (convolution on non-grid, e.g., arbitrary graph) for the task. In this work, we propose to use graph convolutional networks for text classification. We build a single text graph for a corpus based on word co-occurrence and document word relations, then learn a Text Graph Convolutional Network (Text GCN) for the corpus. Our Text GCN is initialized with one-hot representation for word and document, it then jointly learns the embeddings for both words and documents, as supervised by the known class labels for documents. Our experimental results on multiple benchmark datasets demonstrate that a vanilla Text GCN without any external word embeddings or knowledge outperforms state-of-the-art methods for text classification. On the other hand, Text GCN also learns predictive word and document embeddings. In addition, experimental results show that the improvement of Text GCN over state-of-the-art comparison methods become more prominent as we lower the percentage of training data, suggesting the robustness of Text GCN to less training data in text classification.

Topics

class labelsconvolutional neural networkstext classificationtext documentword embeddingattention mechanismelectronic documentnetwork architecturedocument classification

Keywords

Text Graph Convolutional Networksgraph convolutional networkstext classificationdocument classificationnode classificationdocument graphword embeddingsnode embeddings

Table of Contents

  • Introduction
  • Related Work
  • Traditional Text Classification
  • Deep Learning for Text Classification
  • Graph Neural Networks
  • Method
  • Graph Convolutional Networks (GCN)
  • Text Graph Convolutional Networks (Text GCN)
  • Experiment
  • Conclusion and Future Work
  • References

Knowls

  1. Knowl 1 — Heterogeneous Text Graph Construction for a Corpus

    model/method

    To model global word co-occurrence and document-word associations simultaneously, an entire text corpus is represented as a single large heterogeneous graph G=(V,E)G = (V, E).

    The node set VV consists of all unique word nodes and document nodes in the corpus, such that ∣V∣=∣D∣+∣W∣|V| = |D| + |W|, where ∣D∣|D| is the total number of documents (corpus size) and ∣W∣|W| is the vocabulary size of unique words. Every node is initialized with a one-hot representation, corresponding to an identity feature matrix X=I∣V∣×∣V∣∈R∣V∣×∣V∣X = I_{|V| \times |V|} \in \mathbb{R}^{|V| \times |V|}.

    The graph includes two types of edges: word-word edges and document-word edges. The edge weights in the adjacency matrix A∈R∣V∣×∣V∣A \in \mathbb{R}^{|V| \times |V|} are defined as:

    Aij={PMI(i,j)if i,j are words and PMI(i,j)>0TF-IDFijif i is a document and j is a word (or vice versa)1if i=j0otherwiseA_{ij} = \begin{cases} \text{PMI}(i, j) & \text{if } i, j \text{ are words and } \text{PMI}(i, j) > 0 \\ \text{TF-IDF}_{ij} & \text{if } i \text{ is a document and } j \text{ is a word (or vice versa)} \\ 1 & \text{if } i = j \\ 0 & \text{otherwise} \end{cases}

    Point-wise Mutual Information (PMI) is computed over sliding windows of a fixed length moved across all documents in the corpus:

    PMI(i,j)=log⁡p(i,j)p(i)p(j)\text{PMI}(i, j) = \log \frac{p(i, j)}{p(i)p(j)}

    p(i,j)=#W(i,j)#W,p(i)=#W(i)#Wp(i, j) = \frac{\#W(i, j)}{\#W}, \quad p(i) = \frac{\#W(i)}{\#W}

    where #W(i)\#W(i) denotes the number of sliding windows containing word ii, #W(i,j)\#W(i, j) is the number of sliding windows containing both words ii and jj, and #W\#W is the total number of sliding windows in the corpus. Word pairs with PMI(i,j)≤0\text{PMI}(i, j) \le 0 are not connected. Document-word edge weights use term frequency-inverse document frequency (TF-IDF), where term frequency is the count of occurrences of word jj in document ii, and inverse document frequency is the logarithmically scaled inverse fraction of documents containing word jj.

  2. Knowl 2 — Text Graph Convolutional Network Architecture and Training

    model/method

    The Text Graph Convolutional Network (Text GCN) models text classification as a semi-supervised node classification problem on a heterogeneous text graph.

    Given the graph adjacency matrix A∈R∣V∣×∣V∣A \in \mathbb{R}^{|V| \times |V|} with self-loops (Aii=1A_{ii} = 1) and its diagonal degree matrix D∈R∣V∣×∣V∣D \in \mathbb{R}^{|V| \times |V|} where Dii=∑jAijD_{ii} = \sum_j A_{ij}, the normalized symmetric adjacency matrix is defined as A~=D−1/2AD−1/2\tilde{A} = D^{-1/2} A D^{-1/2}.

    A two-layer GCN is applied to the graph. The forward propagation is expressed as:

    Z=softmax(A~ ReLU(A~XW0)W1)Z = \text{softmax}\left(\tilde{A} \, \text{ReLU}\left(\tilde{A} X W_0\right) W_1\right)

    where:

    • X=I∣V∣×∣V∣X = I_{|V| \times |V|} is the identity feature matrix representing one-hot node features,
    • W0∈R∣V∣×kW_0 \in \mathbb{R}^{|V| \times k} is the weight matrix of the first graph convolution layer projecting nodes into a kk-dimensional hidden space,
    • ReLU(x)=max⁡(0,x)\text{ReLU}(x) = \max(0, x) is the activation function,
    • W1∈Rk×FW_1 \in \mathbb{R}^{k \times F} is the weight matrix of the second layer projecting representations into the FF-dimensional class space (FF is the total number of target classes),
    • softmax(xi)=exp⁡(xi)∑f=1Fexp⁡(xf)\text{softmax}(x_i) = \frac{\exp(x_i)}{\sum_{f=1}^F \exp(x_f)} produces the predicted class probability distribution.

    The first-layer node embeddings are given by E1=A~XW0∈R∣V∣×kE_1 = \tilde{A} X W_0 \in \mathbb{R}^{|V| \times k}, and the second-layer embeddings are E2=A~ ReLU(A~XW0)W1∈R∣V∣×FE_2 = \tilde{A} \, \text{ReLU}(\tilde{A} X W_0) W_1 \in \mathbb{R}^{|V| \times F}.

    The model parameters W0W_0 and W1W_1 are trained via gradient descent to minimize the cross-entropy loss over all labeled document nodes YD\mathcal{Y}_D:

    L=−∑d∈YD∑f=1FYdfln⁡Zdf\mathcal{L} = -\sum_{d \in \mathcal{Y}_D} \sum_{f=1}^F Y_{df} \ln Z_{df}

    where Y∈R∣YD∣×FY \in \mathbb{R}^{|\mathcal{Y}_D| \times F} is the binary ground-truth label indicator matrix. A two-layer architecture allows information exchange between documents that are two steps away in the graph (i.e., document →\rightarrow word →\rightarrow document), enabling indirect document-document message passing without requiring direct document citation or similarity edges.

  3. Knowl 3 — Document Classification Performance Across Benchmark Datasets

    empirical result

    The classification performance of Text GCN was evaluated against multiple baseline models across five benchmark corpora (20NG, R8, R52, Ohsumed, and MR). Models were run 10 times, and results are reported as mean test accuracy ±\pm standard deviation.

    Model 20NG R8 R52 Ohsumed MR
    TF-IDF + LR 0.8319±0.00000.8319 \pm 0.0000 0.9374±0.00000.9374 \pm 0.0000 0.8695±0.00000.8695 \pm 0.0000 0.5466±0.00000.5466 \pm 0.0000 0.7459±0.00000.7459 \pm 0.0000
    CNN-rand 0.7693±0.00610.7693 \pm 0.0061 0.9402±0.00570.9402 \pm 0.0057 0.8537±0.00470.8537 \pm 0.0047 0.4387±0.01000.4387 \pm 0.0100 0.7498±0.00700.7498 \pm 0.0070
    CNN-non-static 0.8215±0.00520.8215 \pm 0.0052 0.9571±0.00520.9571 \pm 0.0052 0.8759±0.00480.8759 \pm 0.0048 0.5844±0.01060.5844 \pm 0.0106 0.7775±0.00720.7775 \pm 0.0072
    LSTM 0.6571±0.01520.6571 \pm 0.0152 0.9368±0.00820.9368 \pm 0.0082 0.8554±0.01130.8554 \pm 0.0113 0.4113±0.01170.4113 \pm 0.0117 0.7506±0.00440.7506 \pm 0.0044
    LSTM (pretrain) 0.7543±0.01720.7543 \pm 0.0172 0.9609±0.00190.9609 \pm 0.0019 0.9048±0.00860.9048 \pm 0.0086 0.5110±0.01500.5110 \pm 0.0150 0.7733±0.00890.7733 \pm 0.0089
    Bi-LSTM 0.7318±0.01850.7318 \pm 0.0185 0.9631±0.00330.9631 \pm 0.0033 0.9054±0.00910.9054 \pm 0.0091 0.4927±0.01070.4927 \pm 0.0107 0.7768±0.00860.7768 \pm 0.0086
    PV-DBOW 0.7436±0.00180.7436 \pm 0.0018 0.8587±0.00100.8587 \pm 0.0010 0.7829±0.00110.7829 \pm 0.0011 0.4665±0.00190.4665 \pm 0.0019 0.6109±0.00100.6109 \pm 0.0010
    PV-DM 0.5114±0.00220.5114 \pm 0.0022 0.5207±0.00040.5207 \pm 0.0004 0.4492±0.00050.4492 \pm 0.0005 0.2950±0.00070.2950 \pm 0.0007 0.5947±0.00380.5947 \pm 0.0038
    PTE 0.7674±0.00290.7674 \pm 0.0029 0.9669±0.00130.9669 \pm 0.0013 0.9071±0.00140.9071 \pm 0.0014 0.5358±0.00290.5358 \pm 0.0029 0.7023±0.00360.7023 \pm 0.0036
    fastText 0.7938±0.00300.7938 \pm 0.0030 0.9613±0.00210.9613 \pm 0.0021 0.9281±0.00090.9281 \pm 0.0009 0.5770±0.00490.5770 \pm 0.0049 0.7514±0.00200.7514 \pm 0.0020
    fastText (bigrams) 0.7967±0.00290.7967 \pm 0.0029 0.9474±0.00110.9474 \pm 0.0011 0.9099±0.00050.9099 \pm 0.0005 0.5569±0.00390.5569 \pm 0.0039 0.7624±0.00120.7624 \pm 0.0012
    SWEM 0.8516±0.00290.8516 \pm 0.0029 0.9532±0.00260.9532 \pm 0.0026 0.9294±0.00240.9294 \pm 0.0024 0.6312±0.00550.6312 \pm 0.0055 0.7665±0.00630.7665 \pm 0.0063
    LEAM 0.8191±0.00240.8191 \pm 0.0024 0.9331±0.00240.9331 \pm 0.0024 0.9184±0.00230.9184 \pm 0.0023 0.5858±0.00790.5858 \pm 0.0079 0.7695±0.00450.7695 \pm 0.0045
    Graph-CNN-C 0.8142±0.00320.8142 \pm 0.0032 0.9699±0.00120.9699 \pm 0.0012 0.9275±0.00220.9275 \pm 0.0022 0.6386±0.00530.6386 \pm 0.0053 0.7722±0.00270.7722 \pm 0.0027
    Graph-CNN-S – 0.9680±0.00200.9680 \pm 0.0020 0.9274±0.00240.9274 \pm 0.0024 0.6282±0.00370.6282 \pm 0.0037 0.7699±0.00140.7699 \pm 0.0014
    Graph-CNN-F – 0.9689±0.00060.9689 \pm 0.0006 0.9320±0.00040.9320 \pm 0.0004 0.6304±0.00770.6304 \pm 0.0077 0.7674±0.00210.7674 \pm 0.0021
    Text GCN 0.8634±0.0009\mathbf{0.8634 \pm 0.0009} 0.9707±0.0010\mathbf{0.9707 \pm 0.0010} 0.9356±0.0018\mathbf{0.9356 \pm 0.0018} 0.6836±0.0056\mathbf{0.6836 \pm 0.0056} 0.7674±0.00200.7674 \pm 0.0020

    Text GCN significantly outperforms all baseline methods on 20NG, R8, R52, and Ohsumed (p<0.05p < 0.05 under Student's tt-test) without using pre-trained word embeddings or external knowledge sources. On the Movie Review (MR) dataset, which consists of very short single sentences, CNN-non-static achieves the highest accuracy (0.77750.7775), while Text GCN achieves 0.76740.7674. This is because sentiment classification in short texts depends heavily on local word sequence and word order, and short texts produce fewer graph edges for message passing.

  4. Knowl 4 — Experimental Corpus Statistics and Text GCN Hyperparameters

    experimental setup

    Experiments are conducted across five standard text classification benchmarks:

    Dataset # Docs # Training # Test # Words # Nodes # Classes Average Length
    20NG 18,846 11,314 7,532 42,757 61,603 20 221.26
    R8 7,674 5,485 2,189 7,688 15,362 8 65.72
    R52 9,100 6,532 2,568 8,892 17,992 52 69.82
    Ohsumed 7,400 3,357 4,043 14,157 21,557 23 135.82
    MR 10,662 7,108 3,554 18,764 29,426 2 20.39

    Preprocessing consists of text cleaning, tokenization, removing stop words using NLTK, and removing low-frequency words occurring fewer than 5 times in 20NG, R8, R52, and Ohsumed. For MR, no words are removed due to the short document length.

    For Text GCN training:

    • First-layer hidden representation dimension k=200k = 200,
    • Sliding window size for PMI calculation is 20,
    • Learning rate is 0.020.02,
    • Dropout rate is 0.50.5,
    • L2L_2 weight decay is 00,
    • Optimization is performed using Adam for a maximum of 200 epochs,
    • 10%10\% of the training set is randomly sampled as a validation set for early stopping if validation loss does not decrease for 10 consecutive epochs.

    For baseline models utilizing pre-trained embeddings, 300-dimensional GloVe vectors are used.

  5. Knowl 5 — Classification Robustness under Low Labeled Data Proportions

    empirical result

    When evaluated on training data subsets comprising 1%1\%, 5%5\%, 10%10\%, and 20%20\% of the original training sets of 20NG and R8, Text GCN demonstrates strong label propagation capabilities:

    • On 20NG, Text GCN achieves a test accuracy of 0.8063±0.00250.8063 \pm 0.0025 with only 20%20\% of labeled training documents, outperforming several baseline models trained on 100%100\% of the training set.
    • On R8, Text GCN achieves a test accuracy of 0.8830±0.00270.8830 \pm 0.0027 with only 1%1\% of labeled training documents, which also exceeds multiple baselines trained on the entire training set.

    The performance advantage of Text GCN over baseline models (including CNN-non-static, LSTM with pre-training, Graph-CNN-C, and TF-IDF + LR) becomes substantially more pronounced as the proportion of labeled training data is reduced. This indicates that GCN effectively propagates label signals across the heterogeneous graph via intermediate word nodes.

  6. Knowl 6 — Hyperparameter Sensitivity to Window Size and Embedding Dimension

    empirical result

    The sensitivity of Text GCN performance to the sliding window size and the first-layer hidden embedding dimension was evaluated on the R8 and MR datasets:

    • Sliding Window Size: When tested across window sizes ranging from 5 to 30, test accuracy initially increases as the window size grows, but plateaus once the window size exceeds 15. A window size that is too small fails to capture sufficient global word co-occurrence statistics, whereas an excessively large window size adds edges between words that lack close semantic relationships.
    • First-Layer Embedding Dimension: When tested across embedding dimensions ranging from 50 to 300, test accuracy increases up to dimension 200 and then levels off. Lower dimensions limit the capacity to propagate label information throughout the graph, while higher dimensions increase computational and training time without providing accuracy gains.

  7. Knowl 7 — Interpretability and Semantic Clustering of Learned Word and Document Embeddings

    empirical result

    Text GCN jointly learns representations for both document nodes and word nodes in the same coordinate spaces:

    • Document Embeddings: Visualizing test document embeddings using t-SNE demonstrates that the 20-dimensional second-layer document representations in Text GCN form tighter, more distinctly separated class clusters than the 200-dimensional first-layer embeddings or embeddings obtained from PV-DBOW and PTE.
    • Word Embeddings: Visualizing the 20-dimensional second-layer word embeddings on 20NG (assigning each word's label as the dimension index having the largest scalar value) shows that words with identical labels cluster together by topic.

    Examining the top 10 words with the highest second-layer values for selected classes demonstrates strong semantic correspondence with the respective categories:

    • comp.graphics: jpeg, graphics, image, gif, 3d, images, rayshade, polygon, pov, viewer
    • sci.space: space, orbit, shuttle, launch, moon, prb, spacecraft, solar, mission, alaska
    • sci.med: candida, geb, disease, patients, yeast, msg, vitamin, syndrome, infection, gordon
    • rec.autos: car, cars, v12, callison, engine, toyota, nissan, v8, mustang, eliot

  8. Knowl 8 — Transductive Nature and Inductive Inference Limitation of Text GCN

    limitation

    Text GCN is an inherently transductive model: all test document nodes (without class labels) must be present in the text graph during graph construction and model training to establish edges with word nodes and participate in neighborhood aggregation.

    As a consequence, Text GCN cannot generate representations or make predictions online for newly arrived, unseen test documents without rebuilding the corpus graph and retraining the model or employing inductive graph neural network frameworks.

Coverage note — None was omitted; all primary architectural components, graph construction procedures, empirical benchmark evaluations, low-resource label tests, parameter sensitivity analyses, interpretability findings, and limitations were extracted into knowls.

References

  1. 1.Aggarwal, C. C., and Zhai, C. 2012. A survey of text classification algorithms. In Mining text data. Springer. 163–222.
  2. 2.Bastings, J.; Titov, I.; Aziz, W.; Marcheggiani, D.; and Simaan, K. 2017. Graph convolutional encoders for syntax-aware neural machine translation. In EMNLP, 1957–1967.
  3. 3.Battaglia, P. W.; Hamrick, J. B.; Bapst, V.; Sanchez-Gonzalez, A.; Zambaldi, V.; Malinowski, M.; Tacchetti, A.; Raposo, D.; Santoro, A.; Faulkner, R.; et al. 2018. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261.
  4. 4.Bruna, J.; Zaremba, W.; Szlam, A.; and LeCun, Y. 2014. Spectral networks and locally connected networks on graphs. In ICLR.
  5. 5.Cai, H.; Zheng, V. W.; and Chang, K. 2018. A comprehensive survey of graph embedding: problems, techniques and applications. IEEE Transactions on Knowledge and Data Engineering 30(9):1616–1637.
  6. 6.Chen, J.; Ma, T.; and Xiao, C. 2018. Fastgcn: Fast learning with graph convolutional networks via importance sampling. In ICLR.
  7. 7.Chenthamarakshan, V.; Melville, P.; Sindhwani, V.; and Lawrence, R. D. 2011. Concept labeling: Building text classifiers with minimal supervision. In IJCAI.
  8. 8.Conneau, A.; Schwenk, H.; Barrault, L.; and Lecun, Y. 2017. Very deep convolutional networks for text classification. In EACL.
  9. 9.Defferrard, M.; Bresson, X.; and Vandergheynst, P. 2016. Convolutional neural networks on graphs with fast localized spectral filtering. In NIPS, 3844–3852.
  10. 10.Hamilton, W.; Ying, Z.; and Leskovec, J. 2017. Inductive representation learning on large graphs. In NIPS, 1024–1034.
  11. 11.Henaff, M.; Bruna, J.; and LeCun, Y. 2015. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163.
  12. 12.Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation 9(8):1735–1780.
  13. 13.Joulin, A.; Grave, E.; Bojanowski, P.; and Mikolov, T. 2017. Bag of tricks for efficient text classification. In EACL, 427–431. Association for Computational Linguistics.
  14. 14.Kim, Y. 2014. Convolutional neural networks for sentence classification. In EMNLP, 1746–1751.
  15. 15.Kingma, D., and Ba, J. 2015. Adam: A method for stochastic optimization. In ICLR.
  16. 16.Kipf, T. N., and Welling, M. 2017. Semi-supervised classification with graph convolutional networks. In ICLR.
  17. 17.Le, Q., and Mikolov, T. 2014. Distributed representations of sentences and documents. In ICML, 1188–1196.
  18. 18.Li, Q.; Han, Z.; and Wu, X. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. In AAAI.
  19. 19.Li, Y.; Jin, R.; and Luo, Y. 2018. Classifying relations in clinical narratives using segment graph convolutional and recurrent neural networks (seg-gcrns). Journal of the American Medical Informatics Association DOI: 10.1093/jamia/ocy157.
  20. 20.Liu, P.; Qiu, X.; and Huang, X. 2016. Recurrent neural network for text classification with multi-task learning. In IJCAI, 2873–2879. AAAI Press.
  21. 21.Luo, Y.; Sohani, A. R.; Hochberg, E. P.; and Szolovits, P. 2014. Automatic lymphoma classification with sentence subgraph mining from pathology reports. Journal of the American Medical Informatics Association 21(5):824–832.
  22. 22.Luo, Y.; Xin, Y.; Hochberg, E.; Joshi, R.; Uzuner, O.; and Szolovits, P. 2015. Subgraph augmented non-negative tensor factorization (santf) for modeling clinical narrative text. Journal of the American Medical Informatics Association 22(5):1009–1019.
  23. 23.Luo, Y.; Uzuner, Ö.; and Szolovits, P. 2016. Bridging semantics and syntax with graph algorithms —state-of-the-art of extracting biomedical relations. Briefings in bioinformatics 18(1):160–178.
  24. 24.Luo, Y. 2017. Recurrent neural networks for classifying relations in clinical notes. Journal of biomedical informatics 72:85–95.
  25. 25.Maaten, L. v. d., and Hinton, G. 2008. Visualizing data using t-sne. JMLR 9(Nov):2579–2605.
  26. 26.Marcheggiani, D., and Titov, I. 2017. Encoding sentences with graph convolutional networks for semantic role labeling. In EMNLP, 1506–1515.
  27. 27.Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In NIPS, 3111–3119.
  28. 28.Pang, B., and Lee, L. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In ACL, 115–124.
  29. 29.Peng, H.; Li, J.; He, Y.; Liu, Y.; Bao, M.; Wang, L.; Song, Y.; and Yang, Q. 2018. Large-scale hierarchical text classification with recursively regularized deep graph-cnn. In WWW, 1063–1072.
  30. 30.Pennington, J.; Socher, R.; and Manning, C. 2014. Glove: Global vectors for word representation. In EMNLP, 1532–1543.
  31. 31.Rousseau, F.; Kiagias, E.; and Vazirgiannis, M. 2015. Text categorization as a graph classification problem. In ACL, volume 1, 1702–1712.
  32. 32.Shen, D.; Wang, G.; Wang, W.; Renqiang Min, M.; Su, Q.; Zhang, Y.; Li, C.; Henao, R.; and Carin, L. 2018. Baseline needs more love: On simple word-embedding-based models and associated pooling mechanisms. In ACL.
  33. 33.Skianis, K.; Rousseau, F.; and Vazirgiannis, M. 2016. Regularizing text categorization with clusters of words. In EMNLP, 1827–1837.
  34. 34.Tai, K. S.; Socher, R.; and Manning, C. D. 2015. Improved semantic representations from tree-structured long short-term memory networks. In ACL, 1556–1566.
  35. 35.Tang, J.; Qu, M.; and Mei, Q. 2015. Pte: Predictive text embedding through large-scale heterogeneous text networks. In KDD, 1165–1174. ACM.
  36. 36.Veličković, P.; Cucurull, G.; Casanova, A.; Romero, A.; Liò, P.; and Bengio, Y. 2018. Graph attention networks. In ICLR.
  37. 37.Wang, S., and Manning, C. D. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In ACL, 90–94. Association for Computational Linguistics.
  38. 38.Wang, Y.; Huang, M.; Zhao, L.; et al. 2016. Attention-based lstm for aspect-level sentiment classification. In EMNLP, 606–615.
  39. 39.Wang, G.; Li, C.; Wang, W.; Zhang, Y.; Shen, D.; Zhang, X.; Henao, R.; and Carin, L. 2018. Joint embedding of words and labels for text classification. In ACL, 2321–2331.
  40. 40.Yang, Z.; Yang, D.; Dyer, C.; He, X.; Smola, A.; and Hovy, E. 2016. Hierarchical attention networks for document classification. In NAACL, 1480–1489.
  41. 41.Zeng, Z.; Deng, Y.; Li, X.; Naumann, T.; and Luo, Y. 2018. Natural language processing for ehr-based computational phenotyping. IEEE/ACM transactions on computational biology and bioinformatics 10.1109/TCBB.2018.2849968.
  42. 42.Zhang, Y.; Liu, Q.; and Song, L. 2018. Sentence-state lstm for text representation. In ACL, 317–327.
  43. 43.Zhang, X.; Zhao, J.; and LeCun, Y. 2015. Character-level convolutional networks for text classification. In NIPS, 649–657.

Citation

MLA
Yao, L., et al. “Graph Convolutional Networks for Text Classification”. arXiv, 2018, http://arxiv.org/abs/1809.05679v3.
APA
Yao, L., Mao, C., & Luo, Y. (2018). Graph Convolutional Networks for Text Classification. arXiv. http://arxiv.org/abs/1809.05679v3
Chicago
Yao, L., C. Mao, and Y. Luo. 2018. “Graph Convolutional Networks for Text Classification”. arXiv. http://arxiv.org/abs/1809.05679v3.
Harvard
Yao, L., Mao, C. and Luo, Y. (2018) “Graph Convolutional Networks for Text Classification”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1809.05679v3.
Vancouver
1. Yao L, Mao C, Luo Y (2018) Graph Convolutional Networks for Text Classification. arXiv

BibTeX

@article{yao2018graph,
  title = {Graph Convolutional Networks for Text Classification},
  author = {Yao, Liang and Mao, Chengsheng and Luo, Yuan},
  year = {2018},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1809.05679v3},
  eprint = {1809.05679}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF