Download
1 / 16
160 likes | 268 Views
Detecting Missing Hyphens in Learner Text. Aoife Cahill * , Martin Chodorow ** , Susanne Wolff * and Nitin Madnani * * Educational Testing Service, 660 Rosedale Road, Princeton, NJ 08541, USA {acahill, swolff, nmadnani}@ets.org
E N D
Detecting Missing Hyphens in Learner Text Aoife Cahill*, Martin Chodorow**, Susanne Wolff* and Nitin Madnani* *Educational Testing Service, 660 Rosedale Road, Princeton, NJ 08541, USA {acahill, swolff, nmadnani}@ets.org **Hunter College and the Graduate Center, City University of New York, NY 10065, USA martin.chodorow@hunter.cuny.edu
Outline • Motivation • Baselines • New Model • Experiments and Results • Conclusion
Motivation • Hyphen errors are infrequent • But are an important consideration for students aiming to improve the overall quality of their writing Dogs are lucky… most of them have built in fur coats! Brrrr! From: http://daughternumberthree.blogspot.com
Motivation • Missing hyphen errors are not all lexical • Schools may have more after school sports. • I went to the dentist after school today. • Language Learner text introduces additional complications • My father like play basketball with me.
Baselines • Baseline 1: Collins Dictionary [5,246] • predicts a missing hyphen between bigrams that appear hyphenated in the dictionary • Baseline 2: Wiki (counts) [1,095] • predicts a missing hyphen between bigrams that occur hyphenated more than 1,000 times in Wikipedia • Baseline 3: Wiki (probs) [673,269] • predicts a missing hyphen between bigrams where the probability of the hyphenated form as estimated from Wikipedia is greater than 0.66
New Model • Logistic Regression Model • assigns a probability to the likelihood of a hyphen occurring between wi and wi+1
Data • Training • Well-edited text (San Jose Mercury News) • Error-corrected data mined from Wikipedia Revisions • Combination • Test • Artificial errors: Brown corpus • Learner text: CLC-FCE corpus, TOEFL/GRE essays
Evaluation on Artificial Errors • Brown Corpus: 24,243 sentences, automatically remove hyphens from 2,072 words • Each system makes a prediction for all bigrams about whether a hyphen should appear between the pair of words • precision: how many of the missing hyphen errors predicted by the system were true errors • recall: how many of the artificially removed hyphens the system detected as errors • f-score: the harmonic mean of precision and recall
Evaluation on Learner Text (1) • CLC-FCE corpus: 173 instances of missing hyphen errors
Evaluation on Learner Text (1) • Some observations: • Very low frequency error (173) • Dominated by one lexical item: make-up • Errors are not independent events
Evaluation on Learner Text (2) • Precision-only manual evaluation • Random sample of 100 errors per system detected in 1,000 student essays • 2 native speaker judgements (0.79)
Evaluation on Learner Text (2) • Native Speaker Judgements (Precision)
Conclusions • Logistic Regression Model for predicting missing hyphens in learner text • Trained on: • A corpus of well-edited text • A corpus of automatically mined corrections • In general, the classifiers outperform the baselines, especially in terms of precision Thanks! Questions? Comments? http://blog.ezinearticles.com
