WikiSplit++: Easy Data Refinement for Split and Rephrase

TL;DR

The paper addresses hallucinations and under-splitting in Split and Rephrase by proposing WikiSplit++—a data refinement pipeline that filters unreliable complex–simple pairs via an NLI classifier and disrupts token similarity by reversing the order of simple sentences. By applying this refinement to WikiSplit and training with a smaller, higher-quality dataset, the authors demonstrate that T5-small can produce more splits with fewer hallucinations, as evidenced by improved entailment ratios and other automatic metrics across multiple benchmark datasets. The method also shows generality to other datasets like MinWikiSplit and BiSECT, suggesting that targeted data refinement can meaningfully improve text-to-text generation tasks beyond this specific domain. Overall, WikiSplit++ offers a practical, scalable approach to cleaner data for Split and Rephrase with tangible gains in output quality and safety.

Abstract

The task of Split and Rephrase, which splits a complex sentence into multiple simple sentences with the same meaning, improves readability and enhances the performance of downstream tasks in natural language processing (NLP). However, while Split and Rephrase can be improved using a text-to-text generation approach that applies encoder-decoder models fine-tuned with a large-scale dataset, it still suffers from hallucinations and under-splitting. To address these issues, this paper presents a simple and strong data refinement approach. Here, we create WikiSplit++ by removing instances in WikiSplit where complex sentences do not entail at least one of the simpler sentences and reversing the order of reference simple sentences. Experimental results show that training with WikiSplit++ leads to better performance than training with WikiSplit, even with fewer training instances. In particular, our approach yields significant gains in the number of splits and the entailment ratio, a proxy for measuring hallucinations.

Figures (2)

• Figure 1: Overview of NLI classification. These are actual examples from WikiSplit.
• Figure 2: Overview of sentence-order reversing. To prevent the model from outputting the sentences as they are, the model is trained with the reverse order of the sentences as the gold labels.