Trommer Colloq Today
Jochen Trommer (Leipzig) will be here this afternoon (3:30pm in the Greenberg room, to be precise) to give a colloquium. The talk is entitled “Simultaneous learning of affix meaning and segmentation” and is based on joint work with Sebastian Bank. Immediately following the colloquium will be a social, so come on out for both!
Every formal description of inflectional systems faces two intertwined analytical problems, the Subsegmentation Problem and the Meaning Assignment Problem. Thus the formal interpretation of the German present singular verb paradigm (e.g. for legen, ‘to put’: 1sg leg-e, 2sg leg-st and 3sg leg-t) depends on the meaning assigned to the involved inflectional suffix(es), but meaning assignment itself presupposes specific decisions on the subsegmentation of affix material: If -st is analyzed as an atomic suffix (e.g., Wunderlich & Fabri 2006), there are three markers -e [+1], -st [+2], -t [+3], but if -st is subanalyzed into -s and -t as in Müller (2006), -t can be assigned the more general interpretation [-1] (with the additional markers -e +1 and -s, [+2]).
In this talk, we show that a unified approach to both problems is possible by outlining a learning algorithm that uses optimal patterns of paradigmatic distribution of potential affixes (Pertsova 2011) as the main criterion for computing morpheme meaning and subsegmentation of affix strings. The central idea is that learners apply local optimization in the sense of the Harmonic Serialism version of Optimality Theory (McCarthy 2010): Every optimization step consists in identifying the affix. with the optimal distribution in a paradigm, assigning a morpheme entry (i.e., a phonological shape coupled with a feature specification) to it, and to “freeze” the substrings corresponding to the newly learned affix in the paradigm for further learning and subanalysis. In subsequent steps the same procedure is iteratively applied to the remaining inflectional strings in the paradigm until all phonological material in the paradigm is exhausted and assigned to morpheme entries. As expected in an optimality-theoretic setting, optimization involves a small set of ranked and violable constraints. Significantly, we will show that different rankings of these constraints result in learning morpheme inventories which are optimal for different assumptions on the optimal means to account for imperfect distribution of affixes.