DAGs

DAG stands for "Directed Acyclic Graph", a way to represent a feature structure.

Each feature has two parts: "content" and "pointer". "Pointer" is used for the index (i.e. a cross-reference to another feature structure); "Content" is used for the value (an atomic value or another feature structure). The point of this is simply to add a uniform way to represent the index information.

Parsing with unification

One approach: parse with the CFG only, and use feature structures as a filter to remove unwanted parses. Inefficient.

Another approach: use unification in the chart parser (Earley parser) we discussed previously to a) enrich the representations in the chart, and b) constrain what subtrees get put in the chart by requiring successful unification.

Note that it's possible to radically simplify the CFG rules by treating category information just like any other feature specification.

Unification and semantics

Feature specifications can contain semantic as well as syntactic and morphological features. Note that items restrict the meanings of their dependents as well as their form:

? Two students elapsed

? Colorless green ideas sleep furiously

? Jill baked the lamb and roasted the ham

This takes us into the area of semantic roles (meaning relations among verbs and their arguments) and selectional restrictions (category constraints placed by verbs on their arguments).