title: BiSECT
type: Split and Rephrase
motivation: Why is the dataset part of GEM?
Table of Contents
[Leave this blank, we autogenerate this section and overwrite content]
Dataset Description
- Homepage: https://github.com/mounicam/BiSECT/
- Repository: https://github.com/mounicam/BiSECT/
- Paper: https://aclanthology.org/2021.emnlp-main.500/
- Points of Contact: Joongwon Kim, Mounica Maddela, Reno Kriz
Dataset and Task Summary
This dataset captures the ‘Split and Rephrase’ task, which involves taking long, complex sentences and splitting them into shorter, simpler, and potentially rephrased meaning-equivalent sentences.
BiSECT was created via bilingual pivoting using subsets of the OPUS dataset (Tiedemann and Nygaard, 2004). It spans multiple domains, from web crawl to government documents. The data released here is in English, but data for other European languages are also available upon request.
Compared to previous resources for this task, the resulting dataset was found to contain examples with higher quality, as well as splits that require more significant modifications.
Why is this dataset part of GEM?
BiSECT is the largest available corpora for the Split and Rephrase task. In addition, it has been shown that BiSECT is of higher quality than previous Split and Rephrase corpora, contains a wider variety of splitting operations, and is also available in four languages.
Languages
BiSECT is available in English (en-US), French, Spanish, German.
Meta Information
Dataset Curators
BiSECT was developed by researchers at the University of Pennsylvania and Georgia Institute of Technology. This work is supported in part by the NSF awards IIS-2055699, ODNI and IARPA via the BETTER program (contract 19051600004), and the DARPA KAIROS Program (contract FA8750-19-2-1004).
Licensing Information
The dataset is not licensed by itself, and the source Opus data consists solely of publicly available parallel corpora.
Citation Information
@inproceedings{kim-etal-2021-bisect,
title = "{B}i{SECT}: Learning to Split and Rephrase Sentences with Bitexts",
author = "Kim, Joongwon and
Maddela, Mounica and
Kriz, Reno and
Xu, Wei and
Callison-Burch, Chris",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.500",
pages = "6193--6209"
}
This work also evaluates on the HSplit-Wiki evaluation set, first introduced in the papers below.
@article{Xu-EtAl:2016:TACL,
author = {Wei Xu and Courtney Napoles and Ellie Pavlick and Quanze Chen and Chris Callison-Burch},
title = {Optimizing Statistical Machine Translation for Text Simplification},
journal = {Transactions of the Association for Computational Linguistics},
volume = {4},
year = {2016},
pages = {401--415}
},
@inproceedings{sulem-etal-2018-bleu,
title = "{BLEU} is Not Suitable for the Evaluation of Text Simplification",
author = "Sulem, Elior and
Abend, Omri and
Rappoport, Ari",
booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
month = oct # "-" # nov,
year = "2018",
address = "Brussels, Belgium",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/D18-1081",
doi = "10.18653/v1/D18-1081",
pages = "738--744"
}
Leaderboard
There is currently no leaderboard for this task.
Dataset Structure
Data Instances
Example of an instance:
{
"gem_id": "bisect-train-0",
"source_sentence": "The report on the visit to Bhutan states that the small community has made the task of coordination less complex and success is manifested in the synchronized programming cycles which now apply to all but one of the agencies ( the World Health Organization ) .",
"target_sentence": "The report on the visit to Bhutan says that the small community has made the coordination work less complex . Success manifests itself in synchronized programming cycles that now apply to all but one organism ( the World Health Organization ) ."
}
Data Fields
The fields are the same across all splits.
gem_id- (string) a unique identifier for the instancesource_sentence- (string) sentence to be simplifiedtarget_sentence- (string) simplified text that was split and rephrased
Data Statistics
| dataset | train | validation | test |
|---|---|---|---|
| BiSECT-en | 928,440 | 9,079 | 583 |
| BiSECT-de | 184,638 | 864 | 735 |
| BiSECT-es | 282,944 | 3,638 | 3,081 |
| BiSECT-fr | 491,035 | 2,400 | 1,036 |
| HSplit | -- | -- | 359 |
| Challenge Set | -- | -- | 1,798 |
Dataset Creation
Curation Rationale
BiSECT was constructed to satisfy the need of a Split and Rephrase corpus that is both large-scale and high-quality. Most previous Split and Rephrase corpora (HSplit-Wiki, Cont-Benchmark, and Wiki-Benchmark) were manually written at a small scale and focused on evaluation, while the one corpus of comparable size, WikiSplit, contains around 25% of pairs contain significant errors. This is because Wikipedia editors are not only trying to split a sentence, but also often simultaneously modifying the sentence for other purposes, which results in changes of the initial meaning.
Communicative Goal
The goal of Split and Rephrase is to break down longer sentences into multiple shorter sentences, which has downstream applications for many NLP tasks, including machine translation and dependency parsing.
Source Data
Initial Data Collection and Normalization
The construction of the BiSECT corpus relies on leveraging the sentence-level alignments from OPUS), a collection of bilingual parallel corpora over many language pairs. Given a target language A, this work extracts all 1-2 and 2-1 sentence alignments from parallel corpora between A and a set of foreign languages B.
Next, the foreign sentences are translated into English using Google Translate's Web API service to obtain sentence alignments between a single long sentence $l$ and two corresponding split sentences $s= (s_1, s_2)$, both in the desired language.
To remove noise, the authors remove pairs where $l$ contains a token with a punctuation after the first two and before the last two alphabetic characters, as well as where $l$ contains more than one unconnected component in its dependency tree, generated via SpaCy.
Who are the source language producers?
Opus corpora are from a variety of sources. The BiSECT English training set contains pairs extracted from five datasets: CCAligned, parallel English-French documents from common crawl; Europarl, an English-French dataset from European Parliament; 10^9 FR-EN, an English-French newswire corpus; ParaCrawl, a multilingual web crawl dataset; and UN, multilingual translated UN documents. The BiSECT English test set contains pairs extracted from two additional datasets: EMEA, an English-French parallel corpus made out of PDF documents from the European Medicines Agency; and JRC-Acquis, a multilingual collection of European Union legislative text. Details about the French, Spanish, and German versions can be found in the paper.
Annotations
Annotation process
The training data was automatically extracted, so no annotators were needed. For the English test set, the authors manually selected 583 high-quality sentence splits from 1000 random source-target pairs from the EMEA and JRC-Acquis corpora.
Who are the annotators?
None.
Personal and Sensitive Information
Since this data is collected from OPUS, all pairs are already in the public domain.
Changes to the Original Dataset for GEM
The original BiSECT training, validation, and test splits are maintained in each language to ensure a fair comparison. Note that the original BiSECT English test set was created by manually selecting 583 high-quality Split and Rephrase instances from 1000 random source-target pairs sampled from the EMEA and JRC-Acquis corpora from OPUS.
As the first English challenge set, we include the HSPLIT-Wiki test set, containing 359 pairs. For each complex sentence, there are four reference splits; To ensure replicability, as reference splits, we again follow the BiSECT paper and present only the references from HSplit2-full.
Special Test Sets
In addition to the two evaluation sets used in the original BiSECT paper, we also introduce a second English challenge set. For this, we initially consider all 7,293 pairs from the EMEA and JRC-Acquis corpora. From there, we classify each pair using the classification algorithm from Section 4.2 of the original BiSECT paper. The three classes are as follows:
- Direct Insertion: when a long sentence l contains two independent clauses and requires only minor changes in order to make a fluent and meaning-preserving split s.
- Changes near Split, when l contains one independent and one dependent clause, but modifications are restricted to the region where l is split.
- Changes across Sentences, where major changes are required throughout l in order to create a fluent split s.
We keep only pairs labeled as Type 3, and after filtering out pairs with significant length differences (signaling potential content addition/deletion), we present a second challenge set of 1,798 pairs.
Considerations for Using the Data
Social Impact of the Dataset
Understanding long and complex sentences is challenging for both humans and NLP models. The BiSECT dataset helps facilitate more research on Split and Rephrase as a task within itself, as well as how it can benefit downstream NLP applications.
Impact on Underserved Communities
The data as provided in GEMv2 is in English, French, Spanish, and German, languages with abundant existing resources. However, the dataset creation process introduced in the original paper provides a framework for leveraging bilingual corpora from any language pair found within OPUS.
Discussion of Biases
The Opus corpora used are from a limited set of relatively formal domains, so it is possible that high performance on the BiSECT test set may not transfer to more informal text.
Other Known Limitations
The creation of English BiSECT relies on translating non-English text back to English. While machine translation systems tend to perform well on high-resource languages, there is still a non-negligible chance that there these systems make errors; through a manual evaluation of a subset of BiSECT, it was found that 15% of pairs contained significant errors, while an additional 22% contained minor adequacy/fluency errors. This problem is exacerbated slightly when creating German BiSECT (22% significant errors, 24% minor errors), and these numbers would likely get larger if lower-resource languages were used.
Getting started with in-depth research on the task
The dataset can be downloaded from the original repository by the authors.
The original BiSECT paper proposes several transformer-based models that can be used as baselines, which also compares against Copy512, an LSTM-based model and the previous state-of-the-art.
The common metric used for automatic evaluation of Split and Rephrase, and sentence simplification more generally is SARI. The BiSECT paper also evaluates using BERTScore. Note that automatic evaluations tend to not correlate well with human judgments, so a human evaluation for quality is generally expected for publication. The original BiSECT paper provides templates for collecting quality annotations from Amazon Mechanical Turk.