Source-linked AI summary
ASSET: A Dataset for Tuning and Evaluation of Sentence Simplification Models with Multiple Rewriting Transformations
Fernando Alva-Manchego, Louis Martin, Antoine Bordes, Carolina Scarton, Benoît Sagot, Lucia Specia
TL;DR
Existing simplification benchmarks emphasize individual transformations, limiting evaluation of models in more realistic multi-transformation settings. The paper introduces ASSET, a crowdsourced multi-reference corpus with varied human rewrites, and finds that it captures simplicity characteristics better while exposing weaknesses in current metrics.
Problem
Existing evaluation datasets focus on transformations such as lexical paraphrasing or splitting, limiting assessment of simplification models across varied rewriting operations.
Method
The paper introduces ASSET, a crowdsourced multi-reference dataset containing human simplifications produced with multiple rewriting transformations.
Results
ASSET simplifications are more abstractive and simpler than those in other evaluation corpora, while neither BLEU nor SARI strongly correlates with simplicity ratings.
Takeaways & Limitations
ASSET provides a benchmark for multi-transformation simplification and motivates better automatic evaluation metrics for this setting.
Takeaways & Limitations
The paper identifies a limitation in current evaluation metrics: popular metrics may be unsuitable when multiple rewriting transformations are performed.
Abstract
from arXiv · showhide
In order to simplify a sentence, human editors perform multiple rewriting transformations: they split it into several shorter sentences, paraphrase words (i.e. replacing complex words or phrases by simpler synonyms), reorder components, and/or delete information deemed unnecessary. Despite these varied range of possible text alterations, current models for automatic sentence simplification are evaluated using datasets that are focused on a single transformation, such as lexical paraphrasing or splitting. This makes it impossible to understand the ability of simplification models in more realistic settings. To alleviate this limitation, this paper introduces ASSET, a new dataset for assessing sentence simplification in English. ASSET is a crowdsourced multi-reference corpus where each simplification was produced by executing several rewriting transformations. Through quantitative and qualitative experiments, we show that simplifications in ASSET are better at capturing characteristics of simplicity when compared to other standard evaluation datasets for the task. Furthermore, we motivate the need for developing better methods for automatic evaluation using ASSET, since we show that current popular metrics may not be suitable when multiple simplification transformations are performed.
1 Introduction
Automatic sentence simplification involves lexical, structural, ordering, and deletion transformations, but evaluation datasets often emphasize only one transformation. ASSET addresses this gap with a crowdsourced dataset designed for multi-transformation evaluation and tuning.
- Sentence simplification can replace complex wording, alter syntax, reorder components, or delete unnecessary information while retaining the main idea.
- TurkCorpus mainly contains lexical paraphrasing, whereas HSplit focuses on sentence splitting, so neither represents the full range of simplification transformations.
- These evaluation-data limitations restrict assessment of models in more abstractive scenarios involving substantial modifications to original sentences.
- ASSET contains 23,590 human simplifications associated with 2,359 original sentences, providing a dataset for tuning and evaluating automatic sentence simplification models.
- Neither BLEU nor SARI shows a strong correlation with simplicity ratings when ASSET simplifications are used as references.
- ASSET’s contributions include a multi-transformation dataset and analysis of dataset characteristics relevant to evaluation benchmarking.
2 Related Work
Prior simplification resources vary in transformation coverage, alignment quality, accessibility, and crowdsourcing methodology. These limitations motivate datasets with richer transformations, multiple references, and reproducible evaluation settings.
- Corpus studies show that human simplification includes dropping, splitting, editing, merging, changing, inserting, reordering, and other transformations beyond lexical replacement.
- Overall, evidence from human simplification studies indicates that sentence structure and lexical content both change during simplification.
- Many English Wikipedia simplification datasets use automatic alignments, but Simple English Wikipedia can retain similar syntactic complexity and leave sentences largely unchanged.
- Newsela provides up to five simplicity levels but its restrictive license prevents public sharing of train/test splits, impeding reproducibility.
- TurkCorpus addresses noisy alignment concerns with eight manual references for each of 2,359 original sentences and supports metrics requiring multiple references.
- Earlier crowdsourcing studies used training or collaborative processes, but some collected data is no longer publicly available.
3 Creating ASSET
ASSET extends TurkCorpus by collecting multiple human simplifications of the same sentences through a crowdsourcing protocol that permits varied rewriting transformations. The resulting corpus offers more references, greater variability, and shorter simplifications than comparison resources.
- Data collection: ASSET reuses TurkCorpus’s 2,359 original sentences while collecting manual simplifications with a richer set of rewriting transformations.
- Data collection: Workers were recruited through Amazon Mechanical Turk and screened using approval, experience, location, and qualification requirements.
- Data collection: The qualification test trained workers with explanations and examples of multiple simplification transformations before annotation access.
- Corpus: Each of the 2,359 original sentences received 10 simplifications, producing 23,590 human simplifications in ASSET.
- Instructions: Workers received examples of lexical paraphrasing, splitting, and compression, while choosing which transformations to apply to each sentence.
- Corpus: ASSET examples show that annotators had more freedom to change sentence structure than annotators in the comparison corpora.
- Corpus: Compared with TurkCorpus and HSplit, ASSET contains more references and variability, more natural sentence splitting, and shorter references on average.
4 Rewriting Transformations in ASSET
ASSET measures simplification through multiple rewriting transformations and provides more varied lexical, compression, and rewriting patterns than TurkCorpus and HSplit.
- Text features: The analysis quantifies sentence splits, compression, lexical paraphrasing, word deletion, addition, reordering, exact matches, lexical complexity, and dependency-tree depth.Features are computed for every simplification and aggregated as histograms and percentages.
- Compression: ASSET has more varied compression levels, whereas TurkCorpus and HSplit cluster near a compression ratio of 1.0.HSplit sometimes exceeds 1.0, likely because splitting can require added words for fluency; ASSET’s variability is consistent with information deletion.
- Lexical and rewriting diversity: ASSET paraphrases inputs more diversely than TurkCorpus and HSplit, whose replace-only distances and word-edit ratios are concentrated near zero.Only 0.4% of ASSET simplifications are exact matches, compared with 16.3% for TurkCorpus and 26.5% for HSplit.
- Lexical and rewriting diversity: ASSET distributes lexical-complexity reductions across multiple levels, while HSplit remains concentrated near a ratio of 1.0.The HSplit pattern reflects predominantly structural rather than lexical simplification.
- Structural transformations: ASSET contains more simplifications that reduce syntactic complexity than TurkCorpus and HSplit, although sentence splitting is uncommon and usually produces two sentences.TurkCorpus has at least one split in only 4.6% of cases; ASSET annotators generally did not split, and most splits involved one split.
5 Rating Simplifications in ASSET
Human preference judgments compare ASSET with TurkCorpus and HSplit on fluency, meaning preservation, and simplicity. Judges generally favored ASSET for fluency and simplicity but favored the more meaning-preserving TurkCorpus references.
- Human judgments: The qualification and rating procedures assessed grammaticality, adequacy, and ease of understanding using crowdsourced workers and quality-control checks.Workers completed qualification tests, and submitted ratings were manually reviewed.
- Evaluation protocol: The preference task compared one ASSET simplification with one TurkCorpus or HSplit simplification for each of 359 test-set originals.Workers could select either simplification or mark the pair as similar.
- Evaluation protocol: Judges evaluated fluency, meaning preservation, and simplicity when comparing the references.The criteria asked which sentence was more fluent, best expressed the original meaning, and was easier to understand.
- Results: Judges generally preferred ASSET simplifications for fluency and simplicity, but preferred TurkCorpus simplifications for meaning preservation.The authors attribute the meaning-preservation preference to TurkCorpus’s mainly lexical replacements and minimal content deletion.
- Results: ASSET and HSplit showed similar preference patterns, with larger differences than the ASSET–TurkCorpus comparison.The authors suggest that syntactic restructuring alone may not be sufficient for a sentence to be judged simpler.
6 Evaluating Evaluation Metrics
The paper evaluates BLEU and SARI against human judgements using ASSET references, finding that their correlations are limited when simplifications involve multiple rewriting operations.
- Experimental setup: The study computes sentence-level Pearson correlations between normalized human ratings and BLEU or SARI using ASSET or TurkCorpus references.Human ratings cover Fluency, Adequacy, and Simplicity-related criteria, while HSplit is excluded because prior work found neither metric correlated with human judgements there.
- Metric correlations: BLEU strongly correlates with Meaning Preservation for both ASSET and TurkCorpus references, but its relationship with Simplicity is inconsistent.BLEU shows no correlation with Simplicity using TurkCorpus and moderate correlation using ASSET.
- Metric correlations: SARI’s correlations are positive but low across Fluency, Adequacy, and Simplicity, with significance only for Simplicity using ASSET references.The metric’s strongest supported relationship in this evaluation is with simplicity when ASSET references are used.
- Metric suitability: SARI may be unsuitable for evaluating automatic simplifications that combine multiple rewrite operations.Prior SARI evaluations used judgements tailored to lexical paraphrasing, unlike the broader judgements collected here.
- Text-feature analysis: Meaning judgements correlate negatively with Levenshtein distance and with the proportions of words added, deleted, and reordered, whereas no conclusion is drawn for Simplicity.These analyses relate human ratings to text features in system simplifications.
7 Conclusion
ASSET is a crowdsourced dataset for tuning and evaluating sentence simplification models, built from simplifications applying multiple rewriting transformations. The authors report that it provides more abstractive and simpler references while motivating improved evaluation metrics for multi-transformation simplification.
- Contribution: ASSET is a crowdsourced dataset for tuning and evaluating sentence simplification models.The conclusion presents ASSET as an alternative to publicly available evaluation datasets focused on one transformation type.
- Dataset properties: Annotators were instructed to apply multiple rewriting transformations, producing more abstractive simplifications than existing evaluation corpora.The paper states that ASSET contains simplifications considered simpler than those in other evaluation corpora.
- Implications: The experiments motivate developing new automatic evaluation metrics for simplification involving multiple rewriting operations.This conclusion follows the paper’s analysis of current metrics with ASSET references.
- Implications: The authors hope ASSET’s multi-transformation features will support models serving audiences such as people with low literacy or cognitive disabilities.This is stated as a future-oriented motivation rather than a demonstrated experimental result.