한국언어과학회

This paper aims at modeling topics from TOEFL essay samples in the TOEFL11 corpus. The TOEFL11 corpus is a collection of 12,100 TOEFL writing samples submitted by test-takers from 11 different countries. The paper applied an unsupervised method (i.e. Latent Dirichlet Allocation or LDA) of clustering texts to written samples, with the aim of automatic modeling of topics. For each of the 11 non-native TOEFL test takers, 1,100 TOEFL essays were transformed to a document-term matrix, and then were fed into the LDA function in the R software. The number of potential topics was set to be 8, which was the same number of prompts the test takers had been given when they took the test. The overall accuracy ranged from 83% to 99% depending on the native language of the test takers. Further analysis needs to be conducted to see how reliably the unsupervised LDA method can be used in automatically classifying written samples to potential topics. Nevertheless, the paper provides an empirical foundation that automatic topic modeling can be applied in an unsupervised way even to the writing sample of English learners. (Sungshin Women’s University)