Source-linked AI summary
For the sake of simplicity: Unsupervised extraction of lexical simplifications from Wikipedia
Mark Yatskar, Bo Pang, Cristian Danescu-Niculescu-Mizil, Lillian Lee
TL;DR
The paper asks how lexical simplifications can be extracted from noisy Wikipedia edit histories. It models edit operations and uses metadata to focus on likely simplifications, finding higher precision than baselines and many high-quality pairs absent from SPLIST.
Problem
The paper addresses extracting lexical simplifications from SimpleEW edits, where revisions include simplifications alongside unrelated changes.
Method
The authors combine a probabilistic edit model for mixed operations with metadata-based filtering and PMI ranking of trusted revisions.
Results
The edit model has higher top-100 precision than SIMPL, while both methods produce many high-quality pairs absent from SPLIST, including 62% and 71% of correct pairs.
Takeaways & Limitations
The two complementary approaches provide a starting point for further exploration and yield examples such as “indigenous” → “native” and “annually” → “every year.”
Takeaways & Limitations
An iterative trusted-comment method was omitted because its many parameters made it difficult to tune.
Abstract
from arXiv · showhide
We report on work in progress on extracting lexical simplifications (e.g., "collaborate" -> "work together"), focusing on utilizing edit histories in Simple English Wikipedia for this task. We consider two main approaches: (1) deriving simplification probabilities via an edit model that accounts for a mixture of different operations, and (2) using metadata to focus on edits that are more likely to be simplification operations. We find our methods to outperform a reasonable baseline and yield many high-quality lexical simplifications not included in an independently-created manually prepared list.
1 Introduction
The paper frames simplicity as a style dimension that can broaden access to information, and proposes learning lexical simplifications from Simple English Wikipedia edits. It contrasts this data-driven focus with general syntactic transformation rules and distinguishes simplification from paraphrasing and machine translation.
- Text simplification can make information available to non-native speakers, children, laypeople, and other broader audiences.
- The authors learn lexical simplifications from SimpleEW edit histories, using ComplexEW edits to filter changes that are not simplifications.SimpleEW is edited by human contributors who enforce simplicity through rewriting.
- Their data-driven approach targets item-specific lexical changes such as “collaborate” → “work together,” which general transformation rules cannot capture.
- Simplification trades off meaning preservation against complexity reduction, unlike paraphrasing, and ComplexEW plus SimpleEW form a noisy comparable rather than clean parallel corpus.
2 Method
The method learns lexical simplifications from Simple English Wikipedia edits while separating simplifications from fixes, no-ops, and spam. It combines a probabilistic edit model with metadata-based filtering and ranks candidate replacements from trusted revisions.
- Edit model: Lexical edit instances are extracted as changes A → a between successive article versions, with C and S denoting ComplexEW and SimpleEW version sequences.The method counts article topics or version sequences rather than individual versions when estimating whether an item is rewritten.
- Edit model: The edit model treats fixes, simplifications, no-ops, and spam as distinct operations, targeting both P(o2 | A) and P(a | A, o2).For the initial work, spam is assigned P(o4) = 0.4, and ComplexEW is assumed to contain only fixes.
- Edit model: The model estimates fix behavior from ComplexEW and rewrite probabilities from SimpleEW, using assumptions about comparable fix rates to derive simplification probabilities.ComplexEW edits provide evidence for fix probabilities, while SimpleEW contains both operation types needed to estimate the conditional rewrite distribution.
- Metadata-based methods: Metadata-based methods use revision comments to identify trusted revisions likely to contain simplifications.The seed set is initialized from comments containing a regular-expression match for “simpl,” based on manual inspection of comments from 700K+ SimpleEW revisions.
- Metadata-based methods: SIMPL ranks lexical edits from trusted revisions by point-wise mutual information, while bootstrapping approaches iteratively expand trusted comments or revisions.The comment-expansion approach was omitted because its many parameters made tuning difficult; the revision-expansion approach found few new simplifications before terminating.
- Metadata-based methods: PMI outperformed raw frequency and conditional probability, while differential weighting was proposed to prevent common fixes from receiving high scores.The proposed weighting would distinguish edit instances within a revision instead of assigning them equal trust.
3 Evaluation8
The evaluation processes large Simple English and Complex English Wikipedia revision histories, compares two proposed systems with random and frequency baselines, and assesses extracted pairs through human judgments and agreement with an independently prepared list. The proposed methods outperform the baselines and produce many high-quality simplifications absent from that list.
- Data and evaluation: ∼1.5M revisions from 81,733 Simple English Wikipedia articles and ∼16M revisions from 19,407 Complex English Wikipedia articles were processed.Only 30% of Simple English revisions involved textual changes.
- Data and evaluation: Lexical edit instances were extracted by aligning adjacent-version sentences with tf-idf and identifying longest differing word-boundary segments.The alignment method was chosen because revisions tended to represent small changes.
- Baselines: RANDOM sampled extracted edits uniformly, while FREQUENT returned the most frequent extracted edits as baselines.
- Results: The edit model achieved higher precision than SIMPL, while both proposed systems yielded many high-quality pairs absent from SPLIST and were complementary.They produced 62% and 71% of the correct pairs not included in SPLIST; the evaluation examined only one simplification per eligible source item.
- Results and future work: The authors present examples including “indigenous” → “native,” “permitted” → “allowed,” and “annually” → “every year,” while identifying comparison with machine-translation and paraphrasing systems as future evaluation.