[Paper Review] Mining Naturally-occurring Corrections and Paraphrases from Wikipedia's Revision History

Aurélien Max, Guillaume Wisniewski|arXiv (Cornell University)|Feb 25, 2022

Wikis in Education and Collaboration10 references55 citations

TL;DR

WiCoPaCo is a large corpus of locally edited rewritings from Wikipedia revisions, enabling study and application in spelling correction and paraphrase generation. It provides in-context modification pairs with rich metadata, starting with a French-focused dataset.

ABSTRACT

Naturally-occurring instances of linguistic phenomena are important both for training and for evaluating automatic processes on text. When available in large quantities, they also prove interesting material for linguistic studies. In this article, we present a new resource built from Wikipedia's revision history, called WiCoPaCo (Wikipedia Correction and Paraphrase Corpus), which contains numerous editings by human contributors, including various corrections and rewritings. We discuss the main motivations for building such a resource, describe how it was built and present initial applications on French.

Motivation & Objective

• Create a large resource of locally occurring rewritings from Wikipedia revisions for NLP applications.
• Enable study and evaluation of spelling corrections, paraphrase generation, and related tasks.
• Provide metadata and context to support supervised learning and evaluation.
• Demonstrate initial uses on French to illustrate data utility.

Proposed method

• Extract local modifications by computing differences between consecutive Wikipedia versions using longest common subsequence.
• Restrict to rewriting spans of at most seven words and normalize text (de-wikification, tokenization).
• Apply handcrafted filtering rules to remove meaning-changing or punctuation-only edits and record full paragraph context.
• Store context, original and modified spans, and Wikipedia identifiers including user information; exclude bot edits.
• Output an XML with unique modification identifiers for annotation and reuse.
• Demonstrate applications with spelling-correction and paraphrase-generation analyses using French data.

Experimental results

Research questions

• RQ1 Can Wikipedia revision history yield large-scale, naturally occurring rewritings suitable for NLP tasks?
• RQ2 What types of local rewritings (spelling corrections, paraphrases, etc.) dominate in the corpus and how can they be characterized?
• RQ3 How effective is WiCoPaCo as a resource for spelling correction and paraphrase generation in practice?
• RQ4 Can the methodology be reproduced for other languages and wikis to build multilingual resources?

Key findings

• WiCoPaCo contains 408,816 modifications in context.
• The resource supports diverse rewrite types, including spelling corrections, paraphrases, and meaning changes.
• For French, the corpus enables spelling-error analysis and morpho-syntactic rewriting pattern study.
• A combination of candidate-generation sources substantially improves spelling-correct candidate coverage (hunspell, frequent edit patterns, and corpus-derived edits).
• In spelling-correction experiments, non-word and real-word errors were extracted at 72,493 and 74,100 instances respectively.
• Spelling-correct candidate evaluation shows the best coverage when combining methods, reaching 96.8% for non-word errors and 92.6% for real-word errors.