Computational Characterizations of Vowel Harmony Patterns and Pathologies

This paper provides a computational analysis of 39 attested vowel harmony patterns in a recent typological analysis (Nevins, 2010), in addition to two unattested harmony patterns which have attracted considerable discussion: ‘majority rules’ (Lombardi, 1999; Baković, 2000) and ‘sour grapes’ (Padgett, 1995). It is shown that these attested patterns, unlike the two unattested ones, are subsequential. We therefore hypothesize that subsequentiality is a universal property of vowel harmony. Furthermore, this property is stronger (i.e. more restrictive) than previously established computational universals. The two unattested patterns have been referred to as ‘pathologies’ (Wilson, 2003) because they are predicted to occur through the interaction of standard constraints within Optimality Theory (OT) (Prince & Smolensky, 2004). The fact that these pathological patterns are unattested in natural languages could either be an accidental gap or due to principled factors. The results of this paper are consistent with the latter; in particular the hypothesis that phonological patterns must be subsequential. This paper relies on finite state grammar (FSG) representations of phonological patterns. There are some advantages to using finite state grammars in analyses of phonology. First, they are adequately expressive. Johnson (1972) and Kaplan & Kay (1994) showed that SPE-style rewrite rules and grammars (Chomsky & Halle, 1968) describe regular relations and since SPE-style grammars are adequately expressive, it follows that FSGs are likewise adequately expressive. Second, operations to manipulate and combine these machines are well understood since FSGs form a fundamental chapter of theoretical computer science (Hopcroft et al., 2001). Third, and perhaps most importantly, it becomes possible to obtain insights with FSGs that are much difficult, if not impossible, to realize in either SPE or OT (Kaplan & Kay, 1994; Riggle, 2004; Heinz, 2009).