TL;DR
The paper introduces SANA (Skeleton-based Autoregressive and Non-Autoregressive generation), a two-stage hybrid framework for table-to-text generation. By combining an autoregressive pointer network to extract an ordered factual "skeleton" with an edit-based non-autoregressive decoder that iteratively refines the text under hard lexical constraints, SANA drastically minimizes hallucinations and boosts attribute coverage—achieving a near-perfect PARENT-T recall of 99.47 on the WikiPerson benchmark.
What Problem Does This Paper Address?
Standard table-to-text models rely on autoregressive sequence-to-sequence architectures (like Transformer encoders-decoders) to generate fluent biographies or summaries from key-value tables. However, these models suffer from two major failure modes:
- Low Coverage: The model frequently skips salient attributes present in the table.
- Poor Faithfulness (Hallucination): Autoregressive exposure bias and free decoding often cause the model to introduce ungrounded facts (e.g., claiming a person is a "singer" when the table only says "actress, model").
Existing two-stage "content-planning" pipelines attempt to mitigate this by planning attributes first, but their second-stage autoregressive surface realizers still drift, hallucinate, or drop planned attributes.
Methods
SANA splits table-to-text generation into two distinct, coupled stages:
[Source Table]
│
▼ (Stage 1: Autoregressive Pointer Net)
[Key Token Skeleton]
│
▼ (Stage 2: Edit-based Non-Autoregressive Transformer Decoder)
[Refined Fluent Text via Deletion / Insertion Operations]
- Stage 1: Skeleton Construction (Autoregressive)
- Table Encoder: Key-value attributes are linearized into 4-tuples $(w_i^j, k_i, p_i^+, p_i^-)$ encoding tokens, keys, and relative position offsets.
- Pointer Network: A Transformer decoder predicts target tokens strictly by pointing (copying) from the linearized table to form an ordered "skeleton" of key factual tokens.
- Stage 2: Surface Realization (Edit-Based Non-Autoregressive)
- Inspired by the Levenshtein Transformer, an edit-based decoder iteratively refines the skeleton using three classifiers: Deletion, Placeholder Insertion, and Token Insertion.
- Full Bidirectional Attention: Because it is non-autoregressive, every token generation decision attends bidirectionally to the entire sequence, keeping the text globally grounded.
- Hard Lexical Constraints: During decoding, deletion of tokens originating from the stage-1 skeleton is explicitly forbidden using a constraint mask. The non-autoregressive model can only insert fluent connecting phrases or delete non-skeleton tokens.
Datasets
The authors evaluated SANA on two English Wikipedia-based table-to-text benchmarks (no new datasets were introduced):
- WikiBio: ~728k table-description pairs. Targets are single-sentence biographies (average length: 26.1 tokens). Known to be very noisy, with ~62% of references containing facts not present in the table.
- WikiPerson: ~310k multi-sentence table-biography pairs (average length: 70.6 tokens). Heuristically filtered to remove sentences containing out-of-table entities, making it cleaner and more information-dense.
Key Insights and Ideas
- Coupling Autoregressive Planning with Non-Autoregressive Realization: Non-autoregressive (NAR) text generation often suffers from the "multi-modality" problem (inconsistent parallel generation across multiple valid phrasings). Feeding an autoregressive skeleton as the initial canvas gives the NAR model strong anchor points, neutralizing the multi-modality problem while preserving NAR's speed and bidirectional conditioning.
- Enforcing Invariance via Hard Constraints: By locking the extracted skeleton tokens into the final text via deletion masking, the model mechanically guarantees that planned facts are never omitted during the fluency-generation step.
- Ablation Validates the Duo: Removing hard constraints drops recall and BLEU, but removing the skeleton stage altogether caused performance to collapse (e.g., WikiPerson BLEU dropping from 25.23 to 19.55), proving that standard NAR decoders cannot generate directly from tables without structural anchoring.
Key Results
SANA outperformed previous end-to-end and two-stage methods across both automatic metrics (BLEU, PARENT, and PARENT-T) and human evaluation:
- Record Coverage and Faithfulness on WikiPerson:
- Achieved a PARENT-T Recall of 99.47 (up from 85.21 in OptimTrans and 91.18 in Content-Plan), and an F1 of 61.34 (vs. 56.10 in OptimTrans).
- Outperformed all baselines in BLEU (25.23 vs. 24.56 for OptimTrans, 25.07 for Content-Plan).
- Robustness to Noisy Data (WikiBio):
- Achieved a state-of-the-art BLEU of 45.78 (beating Bert-to-Bert's 45.62 and Struct-Aware's 44.93) and a PARENT F1 of 55.42 (vs. 53.54 for Bert-to-Bert).
- Human Evaluation (100 WikiPerson samples, 0–4 scale):
- SANA achieved top scores across Fluency (3.46), Coverage (3.72), and Correctness (3.11), whereas standard models exhibited sharp trade-offs between coverage and correctness due to hallucinations.
Limitations and Future Work
- Dependence on Overlapping Surface Forms: The skeleton generation relies on heuristic token matching between tables and reference descriptions during training. For tables requiring heavy numerical reasoning, unit conversion, or abstractive rephrasing, simple token-copying pointer networks will struggle.
- Sensitivity to Stage 1 Errors: While the stage-2 decoder can condition on the source table to recover from incomplete skeletons, false inclusions in the skeleton cannot be deleted due to the enforced hard constraints.
- Inference Speed / Cost: Although NAR decoders can be faster than autoregressive decoders per step, iterative insertion/deletion loops combined with a multi-step pointer network introduce multiple distinct inference passes.
Summary updated