PaperlayerView exact revision on arXiv

Source-linked AI summary

Robust (Controlled) Table-to-Text Generation with Structure-Aware Equivariance Learning

Fei Wang, Zhewei Xu, Pedro Szekely, Muhao Chen

arXiv:2205.03972v1cs.CLcs.AIcs.LG

TL;DR

Prior sequence-to-sequence systems linearize tables, limiting their representation of cell relations and making generation brittle to layout changes. LATTICE adds structure-aware attention and transformation-invariant positional encoding to generative models. It improves T5-based models on ToTTo and HiTab and retains promising performance on harder ToTTo, where previous SOTA systems, including those using transformation-based augmentation, see significant drops.

  • Problem

    Prior sequence-to-sequence systems linearize tables, which does not adequately capture cell relations and is brittle when table layouts change.

    Source: §1 · p. 1, §2.2 · p. 4

  • Method

    LATTICE applies equivariance learning with structure-aware attention over related cells and positional encoding that preserves within-cell positions while making positions among cells invariant.

    Source: §1 · p. 2, §2.2 · p. 4

  • Results

    LATTICE improves T5-based models on ToTTo and HiTab and retains promising performance on harder ToTTo, unlike previous SOTA systems, which see significant drops even with transformation-based augmentation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The experiments support structure-awareness and transformation-invariance as key properties for precise and robust description generation from tabular content.

    Source: §5 · p. 9

    Proper use of the technology requires input corpora to be obtained legally and ethically.

    Source: Ethical Considerations · p. 9

Abstract

from arXiv · show

Controlled table-to-text generation seeks to generate natural language descriptions for highlighted subparts of a table. Previous SOTA systems still employ a sequence-to-sequence generation method, which merely captures the table as a linear structure and is brittle when table layouts change. We seek to go beyond this paradigm by (1) effectively expressing the relations of content pieces in the table, and (2) making our model robust to content-invariant structural transformations. Accordingly, we propose an equivariance learning framework, which encodes tables with a structure-aware self-attention mechanism. This prunes the full self-attention structure into an order-invariant graph attention that captures the connected graph structure of cells belonging to the same row or column, and it differentiates between relevant cells and irrelevant cells from the structural perspective. Our framework also modifies the positional encoding mechanism to preserve the relative position of tokens in the same cell but enforce position invariance among different cells. Our technology is free to be plugged into existing table-to-text generation models, and has improved T5-based models to offer better performance on ToTTo and HiTab. Moreover, on a harder version of ToTTo, we preserve promising performance, while previous SOTA systems, even with transformation-based data augmentation, have seen significant performance drops. Our code is available at https://github.com/luka-group/Lattice.

1 Introduction

Controlled table-to-text generation describes highlighted table subparts, but linearizing tables can obscure cell relations and make generation brittle when layouts change. LATTICE adds structure-awareness and transformation-invariance to generative models, improving T5-based models on ToTTo and HiTab and retaining promising performance on a harder ToTTo version.

1 Introduction

  • Controlled table-to-text generation produces focused descriptions of highlighted table subparts, but prior sequence-to-sequence systems represent tables as linear structures.The paper identifies preserving pretrained models’ generation ability while adding structure-awareness and transformation-invariance as nontrivial.Source: §1 · pp. 1–2
  • Linearizing tables can obscure which cells are related, while equivalent layouts can yield inconsistent generations from T5.A row-wise sequence can place one role beside multiple film names without clarifying which film it belongs to.Source: §1 · p. 2
  • LATTICE uses equivariance learning to incorporate table structure and robustness to content-invariant transformations into pretrained generative models.It prunes attention into an order-invariant graph connecting cells in the same row or column, and adjusts positional encodings to preserve within-cell relative positions while making positions among cells invariant.Source: §1 · p. 2
  • LATTICE improves T5-based models on ToTTo and HiTab and retains promising performance on harder ToTTo, where prior SOTA systems, even with transformation-based augmentation, drop significantly.The paper evaluates robustness to content-invariant structural transformations.Source: §1 · p. 2
  • The paper evaluates the framework on ToTTo and HiTab and tests robustness on a harder ToTTo version focused on content-invariant structural transformations.Its contributions also include specifying structure-awareness and transformation-invariance as properties of precise, robust controlled table-to-text generation.Source: §1 · p. 2

2 Method

The method section defines content-invariant table transformations and describes LATTICE’s structure-aware graph masking and transformation-invariant positional encoding. It also presents layout-agnostic input and data augmentation as alternatives for comparison.

2 Method

  • The section first sets out the transformation preliminaries, then details LATTICE’s graph masking and positional encoding, and finally introduces two alternatives.The following points summarize the preliminaries, core method, and alternative techniques.Source: §2 · p. 2

2.1 Preliminaries

  • Content-invariant transformations, such as transpose and row or column shuffles, change layout while preserving the semantics of the same row or column combinations.The section distinguishes these from content-variant changes that alter table semantics.Source: §2.1 · pp. 2–3

2.2 Transformation-Invariant Graph Masking

  • LATTICE removes attention between structurally unrelated cells and links related cells in the same row or column without ordering those links.Its positional encoding preserves relative positions within a cell or metadata but makes positions among cells invariant to their sequence positions and content-invariant transformations.Source: §2.2 · pp. 3–4

2.3 Alternative Techniques

  • Layout-agnostic input reorders cells without layout information, while data augmentation adds transformed layouts to training; LATTICE outperforms both alternatives in performance and robustness.The layout-agnostic technique loses all table layout information, whereas augmentation exposes the base model to more varied training instances.Source: §2.3 · pp. 4–5

3 Experiments

Experiments evaluate LATTICE on ToTTo and HiTab, then test robustness to content-invariant table transformations and ablate its two mechanisms. LATTICE improves over T5 baselines on both benchmarks and retains performance under layout changes.

3.1 Experimental Settings

  • Experiments use ToTTo and HiTab, reporting BLEU-4 and PARENT on both, plus BLEURT on ToTTo.ToTTo includes overlap and non-overlap subsets based on whether a table appeared during training; HiTab focuses on hierarchical tables.Source: §3.1 · p. 5

3.2 Main Results

  • On ToTTo test, LATTICE (T5-small) gains 2.1 BLEU and 0.8 PARENT points over T5-small; LATTICE (T5-base) gains 1.0 BLEU and 1.7 PARENT points over T5-base.On HiTab, LATTICE also consistently outperforms T5 across model sizes.Source: §3.2 · pp. 6–7, Table 1 · p. 7, Table 2 · p. 7

3.3 Robustness Evaluation

  • 3.4 and 4.5 BLEU points: vanilla T5-small and T5-base, respectively, lose this much under combined row shuffle, column shuffle, and transpose.LATTICE guarantees consistent predictions across these content-invariant transformations and scores 3.3 (T5-small) and 2.4 (T5-base) additional BLEU points over layout-agnostic input on original dev data.Source: §3.3 · pp. 7–8, Table 3 · p. 8

3.4 Ablation Study

  • Structure-aware self-attention improves overall performance by 1.3 BLEU points over fully connected self-attention.Improvements on both seen and unseen table subsets indicate better cell-relation capture rather than table memorization.Source: §3.4 · p. 8, Table 4 · p. 8
  • Transformation-invariant positional encoding adds 0.5 BLEU points, with gains mainly on the overlap subset.The authors attribute this pattern to differing highlighted cells for the same training and development tables, which can make memorized layout information hinder in-domain generalization.Source: §3.4 · p. 8, Table 4 · p. 8

4 Related Work

Prior work often linearizes tables, which can obscure relations among cells and lead to unreliable or hallucinated descriptions. LATTICE addresses limitations in structural modeling and transformation robustness that existing approaches leave unresolved.

4 Related Work

  • Linearized inputs can obscure relations among cells, contributing to unreliable or hallucinated descriptions.A row-wise sequence can leave unclear which film a person is associated with.Source: §4 · pp. 8–9
  • Earlier structural methods add token positions or aggregate row and column information, but do not directly provide pretrained generative Transformers with both structure-awareness and transformation-invariance.LATTICE is presented as addressing this gap.Source: §4 · p. 9
  • Prior equivariance techniques for geometric transformations do not directly apply to Transformer-based generation over structured tabular data.LATTICE instead realizes equivariant intermediate representations for content-invariant table transformations.Source: §4 · p. 9
  • Data augmentation adds training overhead and cannot guarantee true equivariance because its transformation coverage is not exhaustive.LATTICE guarantees equivariance through network design without additional training overhead.Source: §4 · p. 9

5 Conclusion

The paper proposes LATTICE, a structure-aware equivariance learning framework for controlled table-to-text generation. Its experiments support the importance of structure-awareness and transformation-invariance for precise, robust descriptions, and the framework may be extended to related tasks.

5 Conclusion

  • LATTICE is proposed as a structure-aware equivariance learning framework for controlled table-to-text generation.The framework is intended to better use structural information in tables.Source: §5 · p. 9
  • Experiments verify the importance of structure-awareness and transformation-invariance for precise and robust descriptions of tabular content.The paper states that these properties align with how information is organized in tables.Source: §5 · p. 9
  • Future research could extend the framework to other data-to-text, tabular reasoning, retrieval, and textual-tabular pretraining tasks.The paper names these as possible directions.Source: §5 · p. 9

Ethical Considerations

The paper states that the framework does not introduce bias toward specific content, while responsible use depends on the data and requires legally and ethically obtained corpora. Its harder ToTTo development set uses content-invariant transformations and retains the original ground-truth generation.

Ethical Considerations

  • The framework focuses on using structural information and does not introduce bias toward specific content.The paper says beneficial versus harmful use depends mainly on the data.Source: Ethical Considerations · p. 9
  • Responsible use requires input corpora to be legally and ethically obtained.The paper places the distinction between beneficial and harmful use mainly on the data.Source: Ethical Considerations · p. 9
  • The harder ToTTo development set uses content-invariant table transformations while retaining the original dataset's ground-truth generation.The passage states this preserves the original ground truth and introduces no further social bias.Source: Ethical Considerations · pp. 9–10
Loading 2205.03972v1…