Do Rules Rule? Evidence from Malayalam Velar Palatalization Sayantan Mandal 1 Catherine T. Best 2 Anne Cutler 3 1. s.mandal@westernsydney.edu.au 2.c.best@westernsydney.edu.au 3.a.cutler@westernsydney.edu.au MARCS Institute of Brain, Behavior and Development The phonological generativity with which babies create utterances in their language has amazed thinkers for centuries [1, 2]. For speech processing, which requires that hierarchical dependencies and bindings be extracted from linear signals [3], Berent [4] argues phonological productivity is constrained not by naturalness concerns but rather by symbolic algorithms specified over algebraic variables organized into equivalence classes (e.g. IDENTITY: *C1__C2; C1, C2, C3…Cn ϵ [C]). Language specific applications of each rule draw on values derived from phonological inventories. Such computational-representational accounts of cognition [e.g, 5, 6] differ from naturalness accounts in making categorical demands on phonological rule outputs. To test this account, we conducted two Forced-Choice experiments with 15 native Malayalam speakers, addressing single-melody velar gemination after preceding vowels [i, e, a] [7, 8]. Two palatalizing suffixes ([k:uka], [k:ə]) figured in the first experiment, an additional non- palatalizing suffix ([kal]) in the second. A male native speaker recorded [i, e, a]-final nonce-stems conjugated with the suffixes, in both palatalized and unpalatalized forms. Participants listened to matched pairs of palatalized and unpalatalized forms (15 repetitions per vowel per suffix), and indicated their preference using a keyboard. An algebraic computational phonology should impose palatalization across the board with licensed suffixes [k:uka] and [k:ə], and avoid it entirely with unlicensed [kal]. A natural phonology, however, should obey the phonetic implicational laws [9] without exception. Co-articulatory (phonetic) and grammatical accounts of the process can also be contrasted, in that highly unnatural [a]_[kkə] contexts create grammatical demands to trigger palatalization, while more natural [i, e]_[kal] environments would block it. This is because, typologically, palatalization after front vowels [i, e] is prevalent, unlike after [a] [10]. In experiment 1 (Figure 1) we found a significant vowel effect, F(1, 14) = 20.01, p < .01, but no suffix effect or interaction. In planned contrasts, [i, e] induced more triggering than [a], p < .05. Experiment 2 (Figure 2) showed main effects of vowel, F(1, 14) = 24.13, p < .01, and suffix, F(1, 14) = 917.72, p < .01, with a vowel*suffix interaction, F(1, 14) = 7.09, p = .01. Planned contrasts again found [i, e] more effective than [a], p < .01, and no significant difference between [i, e]. Thus, identical performance was found for the two licensed suffixes in each experiment, pinning the source of the interaction in Experiment 2 to unlicensed [kal], for which only sparse accidental palatalization occurred with all vowels. Our results thus conform overall to the licensing pattern for suffixes described by Mohanan [7]. More importantly, however, our data show that a phonological mind can generalize rules to novel forms against naturalness concerns [ also11,12]. Such generativity is attainable only through algebraic mechanisms operating on equivalence classes [13,14]. Further, while our data fail to display any bias for the phonetic laws, the lack of bias is not absolute ([i, e] > [a]). We interpret this as reflecting the double duty of phonology: ensuring infinite productivity from finite contrasts, while optimizing phonetic plausibility to attain efficient transmission. Our results display the interaction between an equivalence class of trigger vowels ([i, e, a]) all members of which undergo an algebraic rule and create palatalized forms with licensed suffixes ([k:uka, k:ə]), but not with the unlicensed [kal]. Such computations in phonology may, however, be sensitive to functional