Moderators: Berthold Crysmann and Stephan Oepen Scribe: Petter Haugereid

Introductory Slides

Default unification

• in LKB

Relational constraints

• difference lists
• list operations member, remove, shuffle
• other data types, sets, general relations

Lineariazaton or Discontinuous Parsing

Type resolution (Feature co-occurrence)

Inter-Constituent Dependencies

The mechanisms for checking chart dependencies (lexical entry that requires a particle + particle) function differently in LKB and PET. PET and LKB should be brought closer toghether. Chart Dependencies are checked in LKB before lexical rules apply. Maybe the chart dependencies should be checked after the lexical rules apply. It was suggested that a subcommittee should have a look at this. (Stephan, Berthold, Ann and Ben)

It was suggested by Emily to have an easy form of manipulating difflist information. According to Ann, not knowing where the end of the list is tricky. According to Berthold, grammar writers already simulate list manipulatins in the grammars.

Discontiuous parsing was brought up by Emily (for example parsing of noun and adjective with same case marking, but situated in different positions in the sentence. According to Stephan, this needs LP constraints. Linear constraints once existed in the generator, Ann doesn't have time to build into the parser, but can help someone to do that. Mike Daniels has worked on something similar -Berthold.

Jesse wanted that 4 years ago, but has designed his grammar not to assert that. The ERG has 16 pairs of rules because of lack of type resolution. The processing would go faster if this doubling of rules could be removed. Ann is not optimistic. There are 2 versions of tfs formalisms: One where everything is resolvable to leaf types (the HPSG formalism). This version is endlessly slow. The other other version is the LKB (Carpenter) typing which is a moderate extension of feature structures.

What we want is a hybrid which in some cases gives full resolution. It requires that one is careful about recursion.

Can this be achieved by enforcing gravity? Ann does not see how it could be implemented. Enforced gravity does not work well now. Multiple inheritance with glb types prevent one from being forced down to the desired type that carries the features.

c := a & b &

• [ F1 t1,
  • F2 t2 ].

d := a & b &

• [ F1 t3,
  • F2 t4 ].

By saying that something is of type a' and b' does not mean that one is forced down to either c' or d' since the system creates a necessary glb type that inherits from a' and b' and has c' and d' as subtypes.

It was discussed whether macros should be (re?)introduced. There was agreement that types would do. According to Berthold, macros are unmaintainable. Ann suggested LexDB as an alternative solution to macros for lexical definitions.

Francis mentioned that there will work with defaults in the japanese grammar.