📄 Language Family Relationship Preserved in Non-native English

Summary

Nagata asks whether traces of a writer's mother tongue in their English are strong enough to recover language family structure outside the Indo-European setting. Earlier work by Nagata and Whittaker (2013) had shown this for European learners. This paper tests two hypotheses on Asian English essays from ICNALE, covering both the outer and expanding circles:

• that family relations survive in all non-native English;
• that their survival depends on proficiency.

Method. Each L1 group is modelled with a trigram model over a mix of function words and POS tags, which reduces topic effects. The groups are then clustered hierarchically by a symmetric divergence between models.

Results.

• The trees pair Chinese with Taiwanese English and Japanese with Korean English. This supports preservation, but only for the expanding circle.
• Splitting essays by proficiency does not erase the relationship, so the proficiency hypothesis is rejected.

Explanation. The author proposes a "probabilistic module": learners carry over L1 preferences among grammatically equivalent options, such as compound noun vs. "NN of NN" or adverb placement. These preferences change slowly because no teacher corrects them.

Why it matters. This matters for native language identification and for grammatical error tools.

Key points

• Data: ICNALE, with 400 native essays plus outer-circle groups (Hong Kong, Pakistan, Philippines, Singapore) and expanding-circle groups (China, Indonesia, Japan, Korea, Thailand, Taiwan). Proficiency is labelled A2, B1_1, B1_2 or B2+, with each level mapped to a TOEIC score band.
• Language models: trigram models with Kneser-Ney smoothing, built with the Kyoto Language Modeling toolkit. N-grams seen fewer than five times were dropped.
• Clustering: agglomerative clustering with average linkage. The distance is a per-token log-likelihood ratio in the style of Juang & Rabiner, averaged over both directions to make it symmetric.
• Tagging: a Penn Treebank-tagset POS tagger trained on manually tagged native and learner text, part of which is released as the Konan-JIEM (KJ) learner corpus.
• Conclusions: the paper revises the hypotheses to I′ (preservation holds in the expanding circle) and II′ (preservation does not depend on proficiency). Supporting evidence comes from "NN of NN" frequencies in ICLE and ICNALE and from sentence-initial vs. sentence-final adverb ratios.

Interesting ideas

• I1. Replace content words with POS tags but keep function words (determiners, prepositions, modals, pronouns) as literal tokens, which strips topic while keeping grammatical habits visible
• I2. Compare corpora by how much worse one group's model predicts another group's text, symmetrised and normalised per token, then cluster the results into a tree
• I3. Look at choices among equally grammatical alternatives (compound vs. "of"-phrase, adverb position, clause order), because writers' preferences there escape correction and persist
• I4. Test whether a signal is robust by splitting each group by proficiency and checking that the subgroups still land next to their own group
• I5. Avoid confusing "outer circle" varieties with L1-influenced English, since institutionalised varieties do not cluster by mother tongue

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): a "topic-blind" token view from crate en could act as a style or dialect fingerprint for detecting learner or non-standard input before tagging decisions are made.
• I2 (learn): the same cross-likelihood distance could compare silver treebanks or witness parsers' output distributions and flag which data sources are near-duplicates or outliers.
• I3 (tree): tracking such optional-variant frequencies over UD trees gives a cheap check that the parser handles each variant consistently, and a signal of which dialect of input it is seeing.
• I4 (learn): a similar split-and-reposition test can serve as a negative control, showing that an evaluation metric reflects real structure rather than the skill level of the data.
• I5 (world): when compiling knowledge about text sources, recording the variety of English as distinct from the writer's L1 would keep regional usage from being mistaken for errors.