ToTTo: A Controlled Table-To-Text Generation Dataset
This dataset provides a benchmark for high-precision conditional text generation, addressing the problem of hallucination in table-to-text tasks for researchers in natural language processing.
The authors introduced ToTTo, a dataset with over 120,000 examples for controlled table-to-text generation, where models produce one-sentence descriptions from Wikipedia tables and highlighted cells, and found that existing methods often generate unsupported phrases.
We present ToTTo, an open-domain English table-to-text dataset with over 120,000 training examples that proposes a controlled generation task: given a Wikipedia table and a set of highlighted table cells, produce a one-sentence description. To obtain generated targets that are natural but also faithful to the source table, we introduce a dataset construction process where annotators directly revise existing candidate sentences from Wikipedia. We present systematic analyses of our dataset and annotation process as well as results achieved by several state-of-the-art baselines. While usually fluent, existing methods often hallucinate phrases that are not supported by the table, suggesting that this dataset can serve as a useful research benchmark for high-precision conditional text generation.