Large language models are remarkably good at being fluent, but they are notoriously bad at being faithful. In the context of data-to-text (D2T) generation—the task of converting structured records like tables, knowledge graph triples, or JSON into natural language—this manifests as "hallucination." A model might describe a stock price as rising when the data shows a decline, or attribute a person’s birthplace to the wrong city.
The core of the research in this field is moving away from "black-box" sequence-to-sequence generation toward systems that explicitly ground linguistic output in the underlying structure of the data.
The Fluency-Faithfulness Gap
Standard neural models, particularly those based on the Transformer architecture, are trained to predict the next most likely token based on linguistic patterns. As Joy Mahapatra observes, scaling up model size often improves readability and informativeness but can actually degrade faithfulness. Larger models have stronger "internal linguistic memories" that can override the specific constraints of the provided input data.
This creates a "hallucination" problem where the output is grammatically perfect but factually untethered. Xunjian Yin and Xiaojun Wan argue that traditional metrics like BLEU fail to capture these errors because they reward stylistic overlap rather than logical accuracy.
Symbolic and Logical Grounding
To bridge this gap, researchers have introduced intermediate reasoning steps that force the model to "think" before it speaks. Instead of a direct mapping from Table $\rightarrow$ Text, systems like $T\ell T$ (Table-to-Logic-to-Text) by Eneko Agirre and Iñigo Alonso introduce an intermediate logical form. By generating a symbolic representation of the table's facts first, the model is constrained by logic before it attempts to generate natural language.
Similarly, the OpAtt (Operation-guided Attention) framework by Chin-Yew Lin and colleagues integrates pre-executed symbolic operations—such as mathematical comparisons or subtractions—directly into the neural network. This ensures that when a model describes a numerical trend, it is mathematically grounded in the source data rather than just guessing a plausible-sounding number.
Bidirectional Consistency and Cycle Training
One of the most intuitive ways to ensure fidelity is to treat the problem as a "round trip." If a model generates a description of a data record, a second model should be able to reconstruct the original data record from that description.
Zhuoer Wang demonstrates that cycle training—training a data-to-text model and a text-to-data model as inverses of each other—drastically reduces factual errors. If the generated text contains a hallucination (e.g., an incorrect date), the inverse model will fail to reconstruct the original data, creating a training signal that penalizes unfaithfulness. This bidirectional constraint ensures the model remains a "faithful messenger" of the source information.
Structural Awareness and Content Planning
Hallucinations often occur because models treat structured data as a "flat" sequence of tokens, losing the hierarchical relationships between entities and attributes.
- Structural Embeddings: Heng Ji and Semih Yavuz use "triple-role" and "tree-level" embeddings to help models understand the specific roles entities play within a knowledge graph.
- Memory and Tracking: The AMG (Attend, Memorize and Generate) framework by Wenting Zhao uses a dynamic memory mechanism to track which "slots" in a table have already been mentioned. This prevents the model from omitting critical data or repeating itself—two common forms of unfaithfulness.
- Decomposed Planning: Rather than generating everything at once, models can be split into a "roadmap" phase (content selection) and a "writing" phase. By explicitly deciding what to say before deciding how to say it, systems can significantly reduce the introduction of extraneous, unsupported claims.
Aligning Representations
Finally, researchers are changing how models "see" data during pre-training. The PRINCE paradigm by Song Xu and Xiaodong He uses "prefix-masked decoding" to force models to learn the underlying factual relationships between entities and attributes, rather than just memorizing linguistic sequences.
Even the format of the input matters: researchers have found that representing data as Python dictionaries or nested lists—formats common in the code-heavy datasets used to train modern LLMs—can boost numerical fidelity more effectively than using natural language templates. This suggests that achieving faithfulness is partly a matter of aligning the "language of data" with the "language of the model."