A Classifier-Based Approach to Preposition and Determiner Error Correction in L2 English

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/C08-1022.pdf

Summary

Non-native writers of English often misuse prepositions and determiners. In a small error-annotated corpus built by the authors, determiners accounted for 17% of errors and prepositions for 12%. Usage rules for these words are hard to state, so De Felice and Pulman learn models of correct usage from well-formed native (L1) text instead. They train a maximum entropy classifier on the British National Corpus (BNC). Each occurrence is encoded as a vector of contextual features extracted automatically:

• syntactic information from a full parse (RASP-style grammatical relations),
• Penn Treebank POS tags in a ±3 window,
• WordNet lexicographer classes,
• named-entity information, and other cues.

On held-out BNC text, the models predict the correct preposition 70.06% of the time and the correct determiner 92.15% of the time. The preposition result is competitive with, or better than, comparable published systems. The classifier is then applied to error detection in learner essays from the Cambridge Learner Corpus (CLC), with encouraging first results. The authors note that the topic and vocabulary gap between the BNC and learner writing limits performance. The work shows that linguistically rich classifiers trained only on native text can support grammar checking for L2 writers.

Key points

• Feature sets contain 18 categories for determiners and 13 for prepositions. Semantic information comes from the 40 broad WordNet lexicographer classes, with no word sense disambiguation.
• The preposition classifier handles nine prepositions, namely the most frequent ones: with, to, on, of, in, from, for, by and at. It is trained on 8,898,359 instances and tested on BNC section J (536,193 items).
• Preposition accuracy is 70.06%, compared with a 26.94% baseline (always choosing of), 64.93% reported by Gamon et al. (2008) and 69.00% by Chodorow et al. (2007). Direct comparison is limited because the systems use different data and preposition sets.
• Per-preposition performance does not track training size neatly:
  • of scores very high (P 74.28%, R 90.47%).
  • by does well despite little data, while for does poorly despite plentiful data.
  • Sense counts from the Oxford English Dictionary do not explain the differences.
  • Wrong guesses skew toward the frequent prepositions in, of and to.
• On learner text, unseen or rare nouns hurt the model. For example, internet occurs only 4 times in training, so missing articles before it are never caught. Learners and the model share at/in/on confusions, but learners mostly confuse for/to while the model tends to confuse to/at.