Detecting L2 Writers’ L1s on the Basis of Their Lexical Styles

An earlier version of this chapter was presented at the Second Language Research Forum in 2004 (Jarvis et al., 2004) at a time when we, the authors, had recently become aware of the field of stylometry (see e.g. Holmes, 1998) but were not yet aware of the broader field of machine learning and automated classification (see e.g. Alpaydin, 2004). We had learned from stylometry that authors tend to exhibit relatively consistent patterns of word choice – patterns often referred to as lexical styles (Youmans, 1990) or informally as wordprints (Hoover, 2003). We had also learned from the stylometric literature that, because of the consistency of authors’ lexical styles, the authorship of an anonymous text can be determined with relatively high levels of success by comparing the lexical style found therein with the lexical styles of known authors who are thought to be possible authors of the anonymous text (e.g. Holmes et al., 2001). Impressed with these findings, we were eager to determine whether the same approach could be applied to the identification of not just the specific identities of authors but their profiles more generally. The particular element of writers’ profiles we were interested in identifying was their native language (L1). Because the successful detection of nonnative writers’ L1s on the basis of their lexical styles would be possible only if learners who shared an L1 also shared reliable lexical-style characteristics that were not shared with learners from other L1 backgrounds (cf. Jarvis, 2000), we saw the prospect of being able to identify individual learners’ L1s on the basis of their patterns of language use as a fundamental potential source of evidence for L1 transfer that had received very little attention in the past (but see Ioup, 1984; Odlin, 1996).