Segmentation and UR Acquisition with UR Constraints
Max Nelson · ScholarWorks@UMassAmherst (University of Massachusetts Amherst) · 2019
This paper presents a model that treats segmentation and underlying representation acquisition as parallel, interacting processes. A probability distribution over mappings from underlying to surface forms is defined us- ing a Maximum Entropy grammar which weights a set of underlying representation constraints (URCs) (Apoussidou, 2007; Pater et al., 2012). URCs are induced from observed surface strings and used to generate candidates. Structural ambiguity arising from the com- parison of segmented outputs to unsegmented surface strings is handled with Expectation Maximization (Dempster et al., 1977; Jarosz, 2013). The model successfully learns a simple voicing assimilation rule and segmentation via correspondences between surface phones and input meanings. The trained grammar is also able to segment novel forms affixed with familiar morphemes.