Evaluating Corpora with Word Lists and Word Difficulty
Brent A. Culligan
Aoyama Gakuin Women’s Junior College
https://doi.org/10.7820/vli.v08.1.Culligan
Download this article (pdf)

Abstract
This study examines the application of an IRT analysis of words on lists including the General Service List (GSL), New General Service List (NGSL), Academic Word List (AWL), New Academic Word List (NAWL), and TOEIC Service List (TSL). By comparing line graphs, density distribution graphs, and boxplots for the average difficulty of each word list to related lists, we can get a visualization of the data’s distribution. Japanese EFL students responded to one or more of 84 Yes/No test forms compiled from 5,880 unique real words and 2,520 nonwords. The real words were analyzed using Winsteps (Linacre, 2005) resulting in IRT estimates for each word. By summing the difficulties of each word, we can calculate the average difficulty of each word list which can then be used to rank the lists. In effect, the process supports the concurrent validity of the lists. The analysis indicates the word family approach results in more difficult word lists. The mean difficulties of the GSL and the BNC_COCA appear to be more divergent and more difficult particularly over the first 4000 words, possibly due to the use of Bauer and Nation’s (1993) Affix Level 6 definition for their compilation. Finally, just as we should expect word lists for beginners to have higher frequency words than subsequent lists, we should also expect them to be easier with more words known to learners. This can be seen with the gradual but marked difference between the different word lists of the NGSL and its supplemental SPs.

Keywords
IRT, word difficulty, corpus validity, measurement, vocabulary testing, Yes/No test

Citation
Culligan, B. A. (2019). Evaluating corpora with word lists and word difficulty. Vocabulary Learning and Instruction, 8 (1), 29–38. https://doi.org/10.7820/vli.v08.1.Culligan