When a modern language model describes a dataset, a table, or a knowledge graph, it faces a fundamental tension between fluency and fidelity. Thanks to massive pre-training, neural architectures are exceptionally good at generating smooth, human-like prose. However, this creative fluency is a double-edged sword. Left unconstrained, models frequently prioritize grammatical elegance over factual accuracy, leading to hallucinations (inventing non-existent facts) or omissions (skipping critical data points).
As Gardent et al. observe, while neural models excel at linguistic production, ensuring semantic adequacy remains a primary hurdle. A model might generate a beautifully structured biography or a flawless summary of a sports game, yet completely misreport the underlying numbers or entity relationships. Understanding why this breakdown happens—and how researchers are fixing it—requires looking closely at how structured data interacts with sequential text generation.
Why Models Hallucinate: The Root Causes
To understand the solutions, we must first diagnose why standard language models fail at data-to-text tasks. The literature points to three systemic vulnerabilities:
1. The Topology Gap (The Linearization Problem)
Language models process information sequentially, token by token. Structured data, however, is inherently non-linear; it exists as tables with rows and columns, or knowledge graphs with nodes and edges. To feed this data into a Transformer, researchers typically "linearize" it into a flat string of text.
This flattening process destroys the underlying data topology. As Xing et al. and Färber et al. demonstrate, when a model treats a structured table or graph as a flat sequence, it loses the explicit "who-is-connected-to-whom" context. Lacking a mechanism to map the text back to the record-field hierarchy, the model relies on statistical correlations from its pre-training to bridge the gaps, inadvertently inventing relationships that do not exist in the source data.
2. Loose Alignment in Training Data
Many data-to-text datasets are scraped from the web by pairing structured records with human-written text (such as matching a Wikidata infobox with a Wikipedia introduction). However, these pairs are rarely perfectly aligned.
Vougiouklis et al. highlight that human-written reference texts often contain extraneous background information missing from the data, or conversely, the data contains facts the human chose to omit. When models are trained on these loosely aligned corpora, they learn that it is acceptable—and even expected—to generate ungrounded text or ignore parts of the input to achieve a natural narrative flow.
3. Semantic and In-Context Interference
Even when utilizing large language models (LLMs) via few-shot prompting, factual integrity is easily compromised. Kasner et al. uncover a phenomenon known as semantic interference: when an LLM is given a few examples of data-to-text transformations in its prompt, it frequently bleeds entities, names, or numerical values from those prompt examples directly into the new output, overriding the actual data it was asked to describe.
Architectural Solutions: Enforcing Grounding
To combat these failures, researchers have shifted away from naive end-to-end generation toward frameworks that explicitly enforce factual grounding.
Decoupling Planning from Realization
One of the most effective ways to prevent a model from losing its way is to separate the generation process into two distinct stages: deciding what to say (content planning) and deciding how to say it (surface realization).
[Structured Data] ──> [Content Planner (Select & Order)] ──> [Text Realizer (Fluency)] ──> [Faithful Text]
Instead of letting a single network handle both tasks simultaneously, Moryossef et al. introduce a symbolic planner that maps out the exact structure and order of information before the neural model begins writing. Similarly, Puduppully utilizes macro planning to organize long-form content into structured paragraph outlines. By forcing the system to commit to a factual blueprint first, the downstream language model is strictly constrained, preventing it from inventing details or skipping requested information.
Enforcing Bidirectional Cycle Consistency
Another intuitive approach treats data-to-text generation as a closed loop. If a model generates a text description from a set of table cells, an inverse model (text-to-data) should be able to extract the exact original cells from that generated text.
Wang et al. leverage this insight by implementing a bidirectional cycle-training framework. If the generation model hallucinates an ungrounded fact or omits a crucial attribute, the inverse extraction model will fail to reconstruct the original data structure, triggering a heavy penalty. This loop forces the generator to remain strictly faithful to the input, making it reliable enough for high-stakes domains.
Look-Ahead and Decoding-Time Verification
Rather than trying to fix text after it has been generated, some frameworks intervene directly during the decoding process.
For instance, Qiu et al. developed a plug-and-play decoding method that performs a "look-ahead" prediction. As the model generates text token by token, the system evaluates the factual validity of the upcoming phrase against the source data before committing to it. If a look-ahead phrase contradicts the source, it is pruned from the search space. In a similar vein, Wang et al. introduce content-matching constraints based on Optimal Transport, mathematically bounding the semantic distance between the generated text tokens and the input table fields.
Neuro-Symbolic and Template Distillation
In domains where the tolerance for factual error is absolute zero, researchers have turned to hybrid neuro-symbolic methods. While pure template-based generators are completely faithful, they are brittle and sound robotic.
To capture the best of both worlds, Zhang et al. introduce a framework that distills the creative fluency of a pre-trained language model into a set of human-auditable, rigid templates. By mapping the model's generation capabilities back into a controlled template structure, they eliminate out-of-domain hallucinations entirely while preserving natural phrasing.
Summary
The evolution of data-to-text generation highlights a critical lesson in machine learning: linguistic fluency does not equal factual accuracy. As demonstrated across various benchmarks—from scientific tables (Guo et al.) to complex charts (Hoque et al.) and free-form question answering (Nan et al.)—unconstrained autoregressive models will prioritize looking natural over being correct.
Achieving true faithfulness requires structural awareness. Whether through explicit content planning, cycle-consistency loops, or decoding-time verification, the most successful systems are those that build a rigid bridge between the multi-dimensional topology of data and the linear flow of human language.
Key Foundational Work
Neural Wikipedian: Generating Textual Summaries from Knowledge Base Triples (2017) — contributed 2 concept elements to this direction.