Source-linked AI summary
Deep Graph Convolutional Encoders for Structured Data to Text Generation
Diego Marcheggiani, Laura Perez-Beltrachini
TL;DR
Neural generation from graph-structured data has commonly relied on sequence models that linearize the input graph. This paper uses a graph convolutional encoder to represent the structure directly, and reports that it outperforms standard LSTM encoders on both evaluated tasks.
Problem
Prior neural graph-to-text approaches use recurrent encoders after a separate graph-linearization step, rather than directly exploiting input structure.
Source: §1 · p. 1Method
Results
On both graph-to-sequence tasks, GCN encoders outperform standard LSTM encoders.
Source: §1 · p. 2Takeaways & Limitations
For WebNLG and SR11Deep, the results support explicitly encoding graph structure rather than encoding a linearized input.
Source: §1 · pp. 1–2, §4 · p. 3
Abstract
from arXiv · showhide
Most previous work on neural text generation from graph-structured data relies on standard sequence-to-sequence methods. These approaches linearise the input graph to be fed to a recurrent neural network. In this paper, we propose an alternative encoder based on graph convolutional networks that directly exploits the input structure. We report results on two graph-to-sequence datasets that empirically show the benefits of explicitly encoding the input graph structure.
1 Introduction
The paper replaces graph linearization followed by recurrent encoding with a graph convolutional encoder that directly represents input structure. Across two graph-to-sequence tasks, the GCN encoders outperform standard LSTM encoders.
1 Introduction
- The paper studies WebNLG generation from RDF graphs and SR11Deep realization from semantic dependency graphs.WebNLG involves potentially many KB relations and subtasks; SR11Deep uses linguistically motivated semantic relations whose number is smaller.Source: §1 · p. 1
- Most prior neural approaches linearize graph inputs for recurrent encoders, rather than directly exploiting the input structure.These approaches build on sequence-to-sequence encoder-decoder architectures.Source: §1 · p. 1
- The proposed encoder relies solely on a GCN to encode source graph structure, without using an LSTM to exploit sequential input information.The authors contrast this with earlier work that used sequential information alongside structure.Source: §1 · p. 1
- On both graph-to-sequence tasks, GCN encoders outperform standard LSTM encoders.The comparison is between explicitly encoding structure with GCNs and encoding a linearized version with LSTMs.Source: §1 · p. 2
2 Graph Convolutional-based Generator
The generator uses a GCN to encode graph structure and an attention-based LSTM decoder to produce text conditioned on that graph. Direction-aware, gated message passing and skip connections support the graph encoder.
Graph Convolutional-based Generator
- The overall model conditions its generated text on a directed, labeled input graph.The graph is the source representation whose content is verbalized in the output.Source: §2 · p. 2
Graph Convolutional Encoder
- The GCN computes node representations from graph structure, explicitly modeling edge labels and directions.The encoder adopts the parametrization of Marcheggiani and Titov (2017).Source: §2 · p. 2
- Each GCN update aggregates neighboring nodes with direction-specific parameters; edge-label embeddings, scalar gates, and nonlinearities shape the updates.Stacking GCN layers lets representations incorporate non-immediate neighbors.Source: §2 · p. 2
Skip Connections
- Residual and dense skip connections help gradients flow through stacked layers, enabling deeper GCN encoders.Residual connections sum layer inputs and outputs; dense connections concatenate them so each layer receives outputs from all preceding layers.Source: §2 · p. 2
Graph Convolutional Encoder
- Without scalar gates, preliminary experiments found poor local minima, especially with several GCN layers.Source: §2 · p. 2
Decoder
- An attention-based LSTM decoder generates one word at a time from prior words and a context attending to the GCN graph representation.The model optimizes negative log likelihood.Source: §2 · p. 3
3 Generation Tasks
The paper instantiates graph-to-text generation for two tasks: WebNLG verbalises RDF triples, while SR11Deep realises semantic dependency graphs. Their input representations retain task-specific structure and features.
3 Generation Tasks
- The section introduces task-specific instantiations of the input graph X for graph-to-text generation.WebNLG graphs are built from RDF triples, whereas SR11Deep inputs are semantic dependency graphs.Source: §3 · p. 3, §3.1 · p. 3, §3.2 · p. 3
3.1 WebNLG Task
- WebNLG generates entity descriptions from related RDF triples and requires lexicalisation, aggregation, referring-expression, and segmentation decisions.Relations become nodes connected to subject and object entities through A0 and A1, giving the encoder a hidden state for each relation and supporting arbitrary numbers of KB relations.Source: §3.1 · p. 3
3.2 SR11Deep Task
- SR11Deep uses semantic dependency graphs to represent target sentences with predicate-argument structures; nodes are lemmas and edges carry NomBank and PropBank labels.The task removes complementizers that, commas, and to-infinitive nodes, while node representations also incorporate morphological and punctuation features.Source: §3.2 · p. 3
4 Experiments
Experiments evaluate graph encoders on WebNLG and SR11Deep, with linearised sequence-to-sequence baselines and task-specific GCN configurations. The WebNLG generator also has an extended variant using pretrained embeddings and copying.
4 Experiments
- WebNLG has 18,102 training and 871 development pairs, with seen-category and unseen-category test sets; evaluation focuses on seen categories.SR11Deep has 39,279 training, 1,034 development, and 2,398 test examples; both datasets exclude pairs with more than 50 target words.Source: §4 · p. 3
- The LSTM baselines consume linearised graphs, using depth-first traversal for SR11Deep and task-specific hidden dimensions.The WebNLG baseline uses 256-dimensional embeddings and hidden units; SR11Deep uses 500-dimensional embeddings and hidden states.Source: §4 · p. 4
- The best WebNLG GCN setup uses four layers with residual connections, while SR11Deep uses seven layers with dense connections.WebNLG uses 256-dimensional embeddings and hidden units; SR11Deep uses 500-dimensional embeddings and hidden units.Source: §4 · p. 4
- WebNLG uses delexicalisation, while SR11Deep anonymises entities after compacting named-entity nodes and tagging them with entity types.The SR11Deep procedure distinguishes same-type entities with numerical suffixes.Source: §4 · p. 4
A GCN-based Generator
- The WebNLG GCNEC variant adds pretrained GloVe embeddings and a copy mechanism to the GCN-based model.It represents multi-word subject and object entities as separate nodes linked by Named Entity edges, using six GCN layers and a single-layer LSTM decoder.Source: A GCN-based Generator · p. 4
- Models are evaluated with BLEU, METEOR, and TER, averaging three runs and reporting standard deviations; test-result tables are provided for both tasks.The tables are identified as WebNLG and SR11Deep test results.Source: A GCN-based Generator · p. 4, Table 1 · p. 4, Table 2 · p. 4
5 Results
Across WebNLG and SR11Deep, GCN encoders outperform LSTM encoders; the section also reports fewer qualitative content errors and ablation evidence favoring skip connections.
5 Results
- GCNs beat LSTMs on both tasks: WebNLG by .009 BLEU points and SR11Deep at .647 BLEU versus .377 BLEU.On SR11Deep, adding linguistic features raised the GCN result to .666 BLEU.Source: §5 · pp. 4–5
- On WebNLG, GCNEC beat PKUWRITER by .047 BLEU points and MELBOURNE by .014, but trailed ADAPT.The reported comparison notes that PKUWRITER used an ensemble of seven models and an additional reinforcement-learning step.Source: §5 · p. 4
- On SR11Deep, STUMBA-D and TBDIL scored higher than the GCN model, using separate syntax-prediction, tree-linearization, and morphology modules.The reported scores were .794 for STUMBA-D and .805 for TBDIL, compared with .666 BLEU for GCN with linguistic features.Source: §5 · p. 5
5.1 Qualitative Analysis of Generated Text
- Both models showed repeated and missing source content, but these errors were less evident with GCNs; LSTM outputs also sometimes hallucinated content.Table 3 provides examples of system outputs.Source: §5.1 · p. 5, Table 3 · p. 5
5.2 Ablation Study
- Skip connections were important; residual and dense connections performed similarly, while dense connections were larger and slightly less accurate.The best GCN model had 4.9M parameters versus 4.3M for the baseline; Table 4 reports BLEU averages and standard deviations over three WebNLG development-set runs.Source: §5.2 · p. 5, Table 4 · p. 5
6 Conclusion
The conclusion reports that GCN encoders benefit structured data-to-text generation relative to LSTM sequential encoders on WebNLG and SR11Deep, and proposes testing other graph representations.
6 Conclusion
- On WebNLG and SR11Deep, explicitly encoding structure with GCNs is beneficial relative to sequential LSTM encoding.The comparison concerns structured data-to-text generation.Source: §6 · p. 5
- The paper proposes applying the approach to Abstract Meaning Representations and scoped semantic representations in future work.The conclusion names AMR and scoped semantic representations as other input graph representations to explore.Source: §6 · p. 5
A.1 Training details
The experiments used OpenNMTpy, a shared training setup, and development-set BLEU for early stopping.
A.1 Training details
- All models used batch size 64, Adam with an initial learning rate of 0.001, and a one-layer LSTM decoder.The implementation used OpenNMTpy.Source: §A.1 · p. 8
- Dropout was applied in both encoder and decoder at rate 0.3.Source: §A.1 · p. 8
- Training stopped early using development-set BLEU, with a maximum of 30 epochs.Source: §A.1 · p. 8
A.2 More example outputs
The appendix points to additional generated-text examples for WebNLG and SR11Deep source graphs.
A.2 More example outputs
- Table 5 provides additional examples of generated texts for source WebNLG and SR11Deep graphs.Source: §A.2 · p. 8, Table 5 · p. 9