A classifier-based approach to preposition and determiner error correction in L2 English Rachele De Felice, Stephen G. Pulman Oxford University Computing Laboratory Coling 2008

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Introduction Prepositions(at, by, for, from, in, of, on, to, and with) Determiners(a, the, and null) I study in Boston but I study at MIT. He is independent of his parents, but dependent on his son. Boys like sport. The boys like sport. she ate an apple. she ate the apple.

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Classifier & Features maximum entropy classifier Classifiers

Classifier & Features Features(determiner) Pick the juiciest apple on the tree.

Classifier & Features Features(preposition) John drove to London.

Classifier & Features Baselines(Prepositions) Always choosing the most frequent option, namely of. Baselines(Determiners) Always choosing the most frequent option, namely null.

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Corpus British National Corpus(BNC) Training Data BNC Testing Data A section of the BNC not used in training, section J.

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Evaluation Prepositions

Evaluation Prepositions

Evaluation Prepositions

Evaluation Prepositions

Evaluation Determiners

Evaluation Determiners

Evaluation Determiners

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Testing the model Corpus Cambridge Learner Corpus (CLC) Training Data Extracting 2523 instances of preposition use from the CLC. (1282 correct, 1241 incorrect)

Testing the model Prepositions

Testing the model System error discussions on Prepositions 1)Ungrammatical 2)Misspelled 3)Annotator's benchmark e.g. I received a beautiful present at my birthday. suggests correction: for annotators: on

Testing the model Determiners Instance typeAccuracy Correct92.2% Incorrect<10%

Testing the model System error discussions on Determiners The Lexical items which are not very frequently seen in the BNC. e.g. I saw it in internet. I booked it on Internet.

Outline  Introduction  Classifier & Features  Corpus  Evaluation  Testing the model  Conclusions

Conclusions Using contextual feature based approach to automatic identification and correction of preposition and determiner errors in L1, which achieve an accuracy of 70.06% and 92.15% respectively. Showing how it can be applied to an error correction task for L2 writing.