Introduction
Data-to-text (d2t) generation aims to transform structured data, such as RDF triplets or slot-value pairs, into fluent natural language. A persistent challenge in this domain is maintaining faithfulness—ensuring the generated text accurately reflects the input data without "hallucinating" extra information or omitting key facts. While high-quality, human-curated faithful datasets are ideal for training, they are expensive and scarce. Most large-scale datasets are "noisy" or unfaithful.
The uFACT (Un-Faithful Alien Corpora Training) method addresses this scarcity-quality trade-off. It demonstrates that faithfulness and semantic consistency can be significantly improved by augmenting a small, high-quality target dataset with large volumes of noisy, "unfaithful" data from entirely different domains, referred to as "alien" corpora.
Methodology
The core of uFACT is a mixed-corpus training strategy using a pre-trained T5-Base model. The approach leverages two main techniques:
- Dataset-Specific Tagging: To manage different data structures across corpora, a unique tag (e.g.,
webnlg:,wikiinfo:, orviggo:) is prepended to each input. This allows the model to learn domain-specific mappings within a single multi-task framework. - The uFACT Recipe: The training follows a "quality-over-quantity" rule for the target domain and a "quantity-over-quality" rule for auxiliary alien data. The model learns general linguistic fluency and data-to-text mapping from the large, noisy alien corpora without compromising the faithfulness of the target domain.
Datasets
The study employs three distinct d2t corpora:
- WebNLG (Target): RDF triplets covering diverse topics like buildings or athletes.
- WikiInfo2Text (Alien): Slot-value pairs extracted from Wikipedia infoboxes.
- ViGGO (Alien): A gaming-related dialogue corpus using dialogue acts and attributes.
Each dataset was prepared in two versions: Faithful [F], filtered for high fact-overlap, and Unfaithful [U], representing the original noisy data.
Results and Discussion
The uFACT approach yields several critical insights into d2t training:
- Target Faithfulness is Mandatory: If the target dataset is noisy (U), performance is poor regardless of auxiliary data. High-quality curation of the target domain is essential.
- Alien Faithfulness is Optional: Adding unfaithful alien data (e.g., configuration FU) often outperforms adding only faithful alien data (FF). The model benefits more from the increased volume and variety of fluent sentences than it is hindered by the noise in the alien facts.
- Regularization and Semantic Mapping: Alien corpora act as a regularizer, preventing overfitting to specific target lexical patterns and encouraging robust semantic mapping.
As shown in the results table, the uFACT-trained model (configuration FU) achieved a state-of-the-art METEOR score of 61.81 on the WebNLG benchmark. Furthermore, the PARENT metric—specifically designed to measure d2t faithfulness—showed a dramatic increase from 44.43 (standard unfaithful training) to 58.54 with the uFACT approach (configuration FUU).
Conclusion
uFACT proves that massive manual curation is not a prerequisite for faithful d2t generation. By maintaining a small, faithful target set and supplementing it with large, noisy alien data, models can achieve superior semantic consistency and fluency. Future work will investigate the sensitivity of this method to domain distance and its scalability to larger language models.
Summary updated