MG Parsing as a Model of Gradient Acceptability in Syntactic Islands

Modeling word and morpheme order in natural language as an efficient trade-off of memory and surprisal.

“…Our measure of memory usage generalizes theories that are based on counting numbers of objects stored in incremental memory (e.g., De Santo, 2020; Frazier, 1985; Gerth, 2015; Gibson, 1998; Graf et al, 2015, 2017; Graf & Marcinek, 2014; Kobele et al, 2013; Miller & Chomsky, 1963; Yngve, 1960). Furthermore, for theories where memory is constrained in its capacity for retrieval rather than storage (e.g., Lewis & Vasishth, 2005; McElree et al, 2003), the information locality bound will still hold.…”

“…However, psycholinguistic theories may differ on whether the entropy of memory H M really is the right measure of memory load, and on whether average surprisal S M really is the right predictor of processing difficulty for humans. Therefore, to establish that our information-theoretic processing model generalizes previous theories, we will establish two links:Our measure of memory usage generalizes theories that are based on counting numbers of objects stored in incremental memory (e.g., De Santo, 2020; Frazier, 1985; Gerth, 2015; Gibson, 1998; Graf et al, 2015, 2017; Graf & Marcinek, 2014; Kobele et al, 2013; Miller & Chomsky, 1963; Yngve, 1960). Furthermore, for theories where memory is constrained in its capacity for retrieval rather than storage (e.g., Lewis & Vasishth, 2005; McElree et al, 2003), the information locality bound will still hold.Our predictor of processing difficulty (i.e., average surprisal) reflects at least a component of the predicted processing difficulty under other theories.…”

“…Other early examples include Miller andChomsky (1963) andFrazier (1985)'s measure of syntactic complexity based on counting the number of local nonterminal nodes. More recently, a line of literature has formulated complexity metrics based on how many nodes are kept in incremental memory for how long during parsing, and used linear or ranked combinations of these metrics to predict acceptability differences in complex embeddings (De Santo, 2020;Gerth, 2015;Graf et al, 2015Graf et al, , 2017Graf & Marcinek, 2014;Kobele et al, 2013;Rambow & Joshi, 1994).…”