BibTex

Provide the BibTex-formatted reference for the dataset. Please use the correct published version (ACL anthology, etc.) instead of google scholar created Bibtex.

@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"
}

Intended Use

What is the intended use of the dataset?

Split and Rephrase.

Splitting Criteria

Describe any criteria for splitting the data, if used. If there are differences between the splits (e.g., if the training annotations are machine-generated and the dev and test ones are created by humans, or if different numbers of annotators contributed to each example), describe them here.

While all training data were derived from subsets of the OPUS corpora, different source subsets were used for training v.s., validation and testing. The training set comprised more web crawl data, whereas development and test sets comprised EMEA and EU texts. Details can be found in the BiSECT paper.

Why is the Dataset in GEM?

What does this dataset contribute toward better generation evaluation and why is it part of GEM?

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.

Difference from other GEM datasets

What else sets this dataset apart from other similar datasets in 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 and contains a wider variety of splitting operations.

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.

Modification Details

For each of these changes, described them in more details and provided the intended purpose of the modification

The original BiSECT training, validation, and test splits are maintained to ensure a fair comparison. Note that the original BiSECT 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 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.

In addition to the two evaluation sets used in the original BiSECT paper, we also introduce a second 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:

1. 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.
2. Changes near Split, when l contains one independent and one dependent clause, but modifications are restricted to the region where l is split.
3. 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.

Pointers to Resources

Getting started with in-depth research on the task. Add relevant pointers to resources that researchers can consult when they want to get started digging deeper into 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.

Proposed Evaluation

List and describe the purpose of the metrics and evaluation methodology (including human evaluation) that the dataset creators used when introducing this task.

Existing automatic metrics, such as BLEU (Papineni et al., 2002) and SAMSA (Sulem et al., 2018), are not optimal for the Split and Rephrase task as they rely on lexical overlap between the output and the target (or source) and underestimate the splitting capability of the models that rephrase often.

As such, the dataset creators focused on BERTScore (Zhang et al., 2020) and SARI (Xu et al., 2016). BERTScore captures meaning preservation and fluency well (Scialom et al., 2021). SARI can provide three separate F1/precision scores that explicitly measure the correctness of inserted, kept and deleted n-grams when compared to both the source and the target. The authors used an extended version of SARI that considers lexical paraphrases of the reference.

Language Producers

What further information do we have on the language producers?

N/A.

Data Preprocessing

How was the text data pre-processed? (Enter N/A if the text was not pre-processed)

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 and two corresponding split sentences, both in the desired language.

The authors further filtered the data in a hybrid fashion.

Filter Criteria

What were the selection criteria?

To remove noise, the authors remove pairs where the single long sentence (l) contains a token with a punctuation after the first two and before the last two alphabetic characters. The authors also removed instances where l contains more than one unconnected component in its dependency tree, generated via SpaCy.

Justification for Using the Data

If not, what is the justification for reusing the data?

Since this data is collected from OPUS, all instances are already in the public domain.

Details on how Dataset Addresses the Needs

Describe how this dataset addresses the needs of underserved communities.

The data as provided in GEMv2 is in English, which is a language with abundant existing resources. However, the original paper also provides Split and Rephrase pairs for French, Spanish, and German, while providing a framework for leveraging bilingual corpora from any language pair found within OPUS.

Potential PII Risk

Considering your answers to the PII part of the Data Curation Section, describe any potential privacy to the data subjects and creators risks when using the dataset.

Since this data is collected from OPUS, all pairs are already in the public domain.

Technical Limitations

Describe any known technical limitations, such as spurrious correlations, train/test overlap, annotation biases, or mis-annotations, and cite the works that first identified these limitations when possible.

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.