A Feedback-Augmented Method for Detecting Errors in the Writing of Learners of English Ryo Nagata et al. Hyogo University of Teacher Education ACL 2006
Objective Detect singular-plural errors in English writing –I ate a lot of chicken. –I ate a lot of chickens.
Approach Learn a decision list to separate mass nouns from count nouns –The paper is made of hemp pulp. –I read the paper. Check if the target noun has the correct form –singular or plural
Decision List Training Corpus British National Corpus EDR Corpus Instance format –She ate fried chicken/mass for dinner Feature –Noun phrase components (e.g. fried) –Context words (e.g. she, ate, for, dinner) Sample decision rules –eat -3 mass –fry np mass –for +3 mass –dinner +3 mass
Ranking Decision Rules Rank by log-likelihood ratio Example
Decision List Feedback Training Corpus Use marked essays by English learners More domain-specific Three ways to use –Add into BNC and EDR corpora Feedback corpus too small to affect p(MC|w c ) –Increase weight of feedback corpus –Increase weight of feedback corpus even more
Increasing the Weight of Feedback Corpus Increase the weight of feedback corpus by statistical confidence
Increasing the Weight of Feedback Corpus Even More Take the log of the general corpus’ confidence
Error Detection Use decision list to determine whether a noun is mass or count Step 1: Mass noun in plural form error Step 2:
Error Detection (Cont.) Step 3:
Testing Corpus 47 essays by Japanese English learners 105 errors identified by professional English marker
Experiment Result DL: decision list FB: add directly fb1: increase weight by confidence fb2: increase weight more
Conclusion Decision list better than rule-based and web-based methods Feedback corpus better than general corpus only