Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -45,3 +45,60 @@ configs:
|
|
45 |
- split: test
|
46 |
path: data/test-*
|
47 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
- split: test
|
46 |
path: data/test-*
|
47 |
---
|
48 |
+
|
49 |
+
# WikiSplit++
|
50 |
+
|
51 |
+
This dataset is the HuggingFace version of WikiSplit++.
|
52 |
+
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.
|
53 |
+
The preprocessed WikiSplit dataset that formed the basis for this can be found [here](https://huggingface.co/datasets/cl-nagoya/wikisplit).
|
54 |
+
|
55 |
+
## Usage
|
56 |
+
|
57 |
+
```python
|
58 |
+
import datasets as ds
|
59 |
+
|
60 |
+
dataset: ds.DatasetDict = ds.load_dataset("cl-nagoya/wikisplit-pp", split="train")
|
61 |
+
|
62 |
+
print(dataset)
|
63 |
+
|
64 |
+
# DatasetDict({
|
65 |
+
# train: Dataset({
|
66 |
+
# features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
|
67 |
+
# num_rows: 504375
|
68 |
+
# })
|
69 |
+
# validation: Dataset({
|
70 |
+
# features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
|
71 |
+
# num_rows: 63065
|
72 |
+
# })
|
73 |
+
# test: Dataset({
|
74 |
+
# features: ['id', 'complex', 'simple_reversed', 'simple_tokenized', 'simple_original', 'entailment_prob', 'split'],
|
75 |
+
# num_rows: 62993
|
76 |
+
# })
|
77 |
+
# })
|
78 |
+
```
|
79 |
+
|
80 |
+
### Data Fields
|
81 |
+
|
82 |
+
- id: The ID of the data (note that it is not compatible with the existing WikiSplit)
|
83 |
+
- complex: A complex sentence
|
84 |
+
- simple_reversed: Simple sentences with their order reversed
|
85 |
+
- simple_tokenized: A list of simple sentences split by [PySBD](https://github.com/nipunsadvilkar/pySBD), not reversed in order (often consists of 2 elements)
|
86 |
+
- simple_original: Simple sentences in their original order
|
87 |
+
- entailment_prob: The average probability that each simple sentence is classified as an entailment according to the complex sentence. [DeBERTa-xxl](https://huggingface.co/microsoft/deberta-v2-xxlarge-mnli) is used for the NLI classification.
|
88 |
+
- split: Indicates which split (train, val, or tune) this data belonged to in the original WikiSplit dataset
|
89 |
+
|
90 |
+
## Paper
|
91 |
+
|
92 |
+
Tsukagoshi et al., [WikiSplit++: Easy Data Refinement for Split and Rephrase](https://arxiv.org/abs/2404.09002), LREC-COLING 2024.
|
93 |
+
|
94 |
+
## Abstract
|
95 |
+
|
96 |
+
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).
|
97 |
+
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.
|
98 |
+
To address these issues, this paper presents a simple and strong data refinement approach.
|
99 |
+
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.
|
100 |
+
Experimental results show that training with WikiSplit++ leads to better performance than training with WikiSplit, even with fewer training instances.
|
101 |
+
In particular, our approach yields significant gains in the number of splits and the entailment ratio, a proxy for measuring hallucinations.
|
102 |
+
|
103 |
+
## License
|
104 |
+
|