📄 Investigating the Features that Affect Cue Usage of Non-native Speakers of English

Summary

The paper asks whether English written for non-native readers uses the connective because differently from native-oriented prose, and what predicts where the cue appears. Its answer can then guide text generation for learners. The authors built two comparable 200,000-word science corpora. SUB-BNC is drawn from the BNC (adult science books). CNNSE consists of English texts for intermediate-level (high-school) readers, taken from books published in China and Japan. Two coders marked "explanation" relations signalled by because, along with nucleus/satellite boundaries and the relation of the larger structure each one sits inside. In native text, because almost always falls in the second span. In the easier learner texts it opens the first span noticeably more often, and the difference is significant. C4.5 decision trees were trained on sentence-level and embedding-structure features and checked against a majority-class baseline. Only the discourse relation of the enclosing structure reliably beat the baseline. This gives NLG systems a simple heuristic for cue placement when writing for non-native readers.

Key points

• Flesch Reading Ease is 47.5 for SUB-BNC (difficult) and 68.7 for CNNSE (easy). Relations were grouped following the Carlson, Marcu and Okurowski (2001) manual into 10 coarse classes.
• 228 because tokens were selected using the 3-way coding of Hirschberg and Litman: explanation / other / not considered.
• Inter-coder agreement was 96.9%, 93% and 96.1% across the three annotated aspects, or 86% overall.
• Second-span placement was 96/104 (92.3%) in SUB-BNC versus 88/124 (71%) in CNNSE (χ² = 16.54, p < 0.001).
• The baseline error is 29%. Feature R (relation of the embedding structure) alone gave 16.2 ± 0.7% error, and an R value of cause, contrast, example or explanation predicts first-span because. Sentence features (tense, voice, length, structure) never beat the baseline. Embedding-only feature sets reached about 21–23% error, and removing R pushed the upper bound above baseline.

Interesting ideas

• I1. Two corpora matched on domain, medium and size but differing in measured readability let a stylistic difference be attributed to the audience rather than the topic.
• I2. When coders sort a cue, they get an explicit "not considered" bucket for awkward constructions (e.g. not because … but because), which keeps those constructions out of the data instead of forcing a label.
• I3. A feature counts as useful only if the upper bound of its 95% confidence interval for cross-validated error falls below the majority-class error. This is a cheap, conservative test on small datasets.
• I4. How a clause is placed depends mostly on the discourse relation of the structure that encloses it, not on local properties like tense or length.

For Mova

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

• I1 (back): Generation could keep separate placement and connective preferences for "plain" versus "native" English, learned from a pair of corpora controlled in this way.
• I2 (learn): Silver-data agreement between witness parsers could likewise use an abstain/excluded class, so unusual constructions are not counted as disagreements or as bad labels.
• I3 (learn): Self-induced repair rules or new features could be accepted only when their CI-bounded error clearly beats a trivial baseline, which acts as a built-in negative control.
• I4 (back): When rendering a cause frame from the meaning graph, the generator could read the parent rhetorical relation to decide whether the because-clause should come first or second.