WikiSplit++

This dataset is the HuggingFace version of WikiSplit++.
WikiSplit++ enhances the original WikiSplit by applying two techniques: filtering through NLI classification and sentence-order reversing, which help to remove noise and reduce hallucinations compared to the original WikiSplit.
The preprocessed WikiSplit dataset that formed the basis for this can be found here.

Dataset Description

• Repository: https://github.com/nttcslab-nlp/wikisplit-pp
• Paper: https://arxiv.org/abs/2404.09002
• Point of Contact: Hayato Tsukagoshi

Usage

import datasets as ds

dataset: ds.DatasetDict = ds.load_dataset("cl-nagoya/wikisplit-pp", split="train")

print(dataset)

# DatasetDict({
#     train: Dataset({
#         features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
#         num_rows: 504375
#     })
#     validation: Dataset({
#         features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
#         num_rows: 63065
#     })
#     test: Dataset({
#         features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
#         num_rows: 62993
#     })
# })

Data Fields

• id: The ID of the data (note that it is not compatible with the existing WikiSplit)
• complex: A complex sentence
• simple_reversed: Simple sentences with their order reversed
• simple_tokenized: A list of simple sentences split by PySBD, not reversed in order (often consists of 2 elements)
• simple_original: Simple sentences in their original order
• entailment_prob: The average probability that each simple sentence is classified as an entailment according to the complex sentence. DeBERTa-xxl is used for the NLI classification.
• split: Indicates which split (train, val, or tune) this data belonged to in the original WikiSplit dataset

Paper

Tsukagoshi et al., WikiSplit++: Easy Data Refinement for Split and Rephrase, LREC-COLING 2024.

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.

License

WikiSplit is distributed under the CC-BY-SA 4.0 license.
This dataset follows suit and is distributed under the CC-BY-SA 4.0 license.