📄 May I check the English of your paper!!!

Summary

This paper describes a Jadavpur University entry to the HOO (Help Our Own) shared task of 2011. The task asks systems to automatically fix English errors in research papers, mostly those written by non-native authors. The authors combine two components. The first is a rule-based layer that uses the Stanford dependency parser and POS tags, together with handmade resources: a preposition list, a verb paradigm table and an adjective/adverb dictionary. Its rules cover prepositions, verb forms, agreement, countability and serial-comma punctuation. The second is a statistical determiner model built with CRF++. It was trained on the HOO training data plus the authors' own published papers, using word, root, POS, number and position features. A set of post-correction rules then refines the CRF output by checking whether a noun was already mentioned earlier in the paragraph. The outputs of all parts are merged. The single submitted run scored modest F-scores of 0.204 for detection, 0.178 for recognition and 0.167 for correction. The work shows a pragmatic hybrid design: hand rules handle the regular error types, and a learned model handles articles, which resist rules.

Key points

• Results from one submitted run: F-score 0.204 for detection, 0.178 for recognition and 0.167 for correction.
• Errors are sorted into coarse groups (wrong form, missing, needs replacing, unnecessary), each crossed with a POS category. Example tags include FT, RT, FV, FD, FY, FJ, RC, RP, UP, CN, CQ, AGV and MT.
• The rule-based model relies on these resources:
  • Stanford dependency parser output, used to find the subject for verb agreement;
  • a manual preposition list and a verb paradigm table;
  • word-position cues to tell adverbs from adjectives;
  • small handcrafted rules for missing prepositions, such as "all" → "all of" and parallel prepositions in "not only … but".
• The determiner model is CRF++ with a trigram setup. Its training marks each determiner plus the two words after it, and also a preceding preposition. Training data was expanded with the authors' own papers.
• The authors admit failures on many syntactic and semantic errors, such as a wrongly used "a". They also note that one error can carry several valid tags.

Interesting ideas

• I1. Two-of-three agreement vote. Countability errors are fixed by majority vote among three signals: the quantifier, the noun's number and the verb. Whichever signal disagrees with the other two is treated as the error, which turns a vague inconsistency into a specific, explainable edit.
• I2. Givenness decides the article. If a noun already appeared earlier in the paragraph, the system inserts "the" or replaces "a" with "the". Discourse history thus drives the article choice, not local context alone.
• I3. Writers have personal error profiles. Listing errors per document showed that each author keeps repeating the same kinds of mistakes. Profiling errors by source can therefore guide where corrections are needed.
• I4. Rules first, statistics for the irregular, rules again after. Hand rules cover the regular phenomena, a sequence labeler takes the hard one (determiners), and a second rule layer cleans up the labeler's output.
• I5. Operation-based error taxonomy. Each error is labeled by an edit type (wrong form, missing, replace, unnecessary) crossed with a POS category. This makes coverage gaps easy to see; for example, the "missing" group was clearly the hardest.

For Mova

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

• I1 (tree): When Number features disagree across a det–noun–verb triple in a UD tree, a majority rule could choose which token's morphology to repair and record the reason for the choice.
• I2 (back): During generation from the world of the text, an entity already introduced should get a definite article, and a new one an indefinite article, directly from the entity registry.
• I3 (learn): Disagreements among witness parsers could be grouped by source or genre, so that repair rules are induced for each recurring error profile.
• I4 (learn): Self-induced repair rules could be applied as a post-pass over gparse or perceptron outputs, targeting known systematic slips instead of retraining for every one.
• I5 (learn): Evaluation reports could break parser and frame errors down by edit type × POS, which shows which repairs are absent and helps design negative controls.