Skip to content
/ TDParse Public

Type-driven syntactic and semantic effectful parsing for NL

Notifications You must be signed in to change notification settings

schar/TDParse

Folders and files

NameName
Last commit message
Last commit date

Latest commit

5197047 · Sep 22, 2022

History

80 Commits
Sep 10, 2022
Sep 2, 2022
Sep 19, 2022
Sep 19, 2022
Sep 19, 2022
Sep 22, 2022
Sep 5, 2022
Sep 22, 2022
Sep 19, 2022
Sep 5, 2022
Sep 1, 2022

Repository files navigation

Type-driven syntactic and semantic effectful parsing for natural language.

$ ghci TDDemo.hs

Handles functors, applicatives, monads, adjoints, and commutative effects. Currently supports Reader for anaphora and context-sensitivity, various Writer's for not-at-issue content and dref introduction, List/Set for nondeterminism and indefiniteness, and Continuations for scope and quantification. Also:

  • Finds all analyses consistent with input, lexical category/type assignments, and type-driven composition rules.

  • No type-shifting! This is an extended type-driven interpreter, in the mold of Klein & Sag 1985 or Heim & Kratzer 1998. We also do not instantiate effectful combinators in the syntax. Even so, useful higher-order constructs like composite functors and composite applicatives (and even some composite monads) arise compositionally.

  • A new treatment of dynamic binding, decomposing the nondeterministic state monad (StateT []) into adjoint Reading and Writing functors, together with nondeterminism. The approach has an intriguing resemblance to algebraic effects, and other operationalizations of free monads.

    On this perspective, the State monad transformer arises compositionally, and (dynamic) binding arises via the counit of the adjunction W -| R.

  • Includes a notion of normal form derivation to limit spurious ambiguity, in most cases (those not having to do with pronoun binding!) exploring exactly one derivation per meaning. Helps greatly with usability, and efficiency.

  • Includes a treatment of scope islands (essentially the one in SC's PhD thesis) and crossover (in terms of a simple restriction on the application of W -| R's counit). Also aids usability and efficiency (as discussed here).

  • Appears pretty efficient, from our informal benchmarking. Seems to handle massive ambiguity with relative aplomb.

Output can be displayed as natural deduction proofs, or as trees annotated with types, rules, and (optionally) semantic values. Here's an example of latex tree output with types and rules:

ghci> outTrees' withDens demoCFG demoLex (hasType (effS T)) "the cat near someone2 saw her"

image

There are a number of routines for viewing output defined in TDPretty.hs.

There is an online version of the software here (ported by DB to Purescript). You can read some notes on the general approach here.

To come/needs done:

  • Integration with a robust/industrial-strength CCG substrate (e.g., OpenCCG)

  • More rigorous benchmarking of soundness and efficiency

  • More HOAS, letting the Haskell compiler do the heavy lifting?

  • Rules for dref introduction, normal form binding derivations. (Sloppy) ellipsis and paycheck pronouns!

About

Type-driven syntactic and semantic effectful parsing for NL

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published