Learning abstract underlying representations from distributional evidence * Ezer Rasin & Roni Katzir Massachusetts Institute of Technology and Tel Aviv University 1. Introduction Human learners have been argued to acquire underlying representations (URs) that are sometimes different from their corresponding surface representation (SR) even without be- ing forced to do so by an alternation. Consider, for example, the case of coalescence in Sanskrit discussed in McCarthy (2005). In Sanskrit, the underlying sequence /ai/ under- goes coalescence and surfaces as the long mid vowel [e:], as in (1). While some long mid vowels have alternants that suggest an underlying /ai/, as in (1), for some long mid vowels the learner will not encounter any alternations supporting a coalescence source, as in the simple constructed example in (2). (1) /tava indra/ [tave:ndra] ‘for you, Indra (voc.)’ (2) [be:] (non-alternating) Given a grammar with coalescence, a non-alternating form such as (2) can be represented in the lexicon in one of two ways. It can be represented either with an identical UR (/be:/), or with the underlying sequence /ai/, as in /bai/, in which case the correct surface form will be derived by the grammar through coalescence. McCarthy (2005) argued that the correct representation within Sanskrit of a form like [be:] is the abstract UR /bai/, despite the absence of supporting evidence from an alternation in this case. In particular, he argued that learners that posit the faithful UR (e.g., the UR /be:/ in the case of the surface form [be:]) converge on an incorrect, over-generating grammar for Sanskrit. We will use the term abstract URs to informally refer to URs such as /bai/ – that is, URs which contain deviations from SRs that are not directly supported by an alternation. Ab- stract URs were occasionally assumed within early generative phonology: as a well-known example, Chomsky & Halle’s (1968) proposed representation of the English word nightin- * We thank Adam Albright, Naomi Feldman, Michael Kenstowicz, Donca Steriade, and the audiences at MIT and NELS 48.