July has been a good month: * Our ACL paper about Wikipedia quality was awarded an SAC highlight (aclanthology.org/2026.acl-lon...) * @colemanhaley.bsky.social joined our lab as a postdoc * Coleman's CoNLL paper about impossible languages won the best paper award (aclanthology…
How Good is Your Wikipedia? Auditing Data Quality for Low-resource and Multilingual NLP aclanthology.org
AI Weekly's analysis
→
- MinHash deduplication removes 28.33% of all non-English Wikipedia articles, mostly from editions known to be dominated by bot-generated content.
- Cited bot-content shares reach 99% for Cebuano, 90% for Waray and 68% for Swedish Wikipedia, per an Alshahrani et al. 2023 estimate.
- Language models trained on the filtered Wikipedia largely match or outperform those trained on the raw dumps, with the biggest gains on lower-quality editions.
Read full analysis →
View on Bluesky ·
♥ 11
↻ 1
↩ 1
·
3 from the directory shared this ·
41d ago
July has been a good month: * Our ACL paper about Wikipedia quality was awarded an SAC highlight (aclanthology.org/2026.acl-lon...) * @colemanhaley.bsky.social joined our lab as a postdoc * Coleman's CoNLL paper about impossible languages won the best paper award (aclanthology…
When transformers learn “impossible” languages, what do they learn? aclanthology.org
AI Weekly's analysis
→
- Janarthan, Haley and Goldwater train GPT-2 style models on perturbed 'impossible' variants of English and probe them beyond perplexity.
- On BLiMP minimal pairs the models show only gradual degradation on impossible languages, mediated by information locality.
- Generation is where the bias lives: the same models produce substantially fewer high-quality sentences at longer lengths.
Read full analysis →
View on Bluesky ·
♥ 11
↻ 1
↩ 1
·
2 from the directory shared this ·
41d ago
