Generating a bilingual lexical corpus using interlanguage normalized Levenshtein distances

Finding large numbers of target items for phonetic and phonological experiments can be a time-consuming and error-prone task. Using freely available tools and data, we have generated a bilingual corpus with the specific aim of investigating the processing and perception of stress in second-language (L2) words. Normalized Levenshtein distances between orthographic and phonemic transcriptions of Brazilian Portuguese (BP) and American English (AmE) translation word pairs were used to automatically generate similar and dissimilar word pairs. Frequency data from corpora were used as a metric of familiarity. To test if these generated metrics correspond to speakers' representations, BP L1 speakers of AmE L2 rated the word pairs on orthographic and phonological similarity, and indicated their familiarity with the English words. Results showed a high correlation between subjective ratings and the computed similarity and familiarity values of the bilingual corpus. We conclude that automatically constructed bilingual corpora such as ours, combined with simple string similarity metrics, are a valid and useful tool for experimental research into L2 (word stress). Key-words: Phonetics, Corpus Linguistics, Psycholinguistics, normalized Levenshtein distances, L2 segmental categorization, L2 word stress