CCGlab is a tool to explore with Combinatory Categorial Grammar (CCG), from quick check of a linguistic analysis in CCG, to large-scale development of grammars and models.

-Cem Bozsahin

CCGlab grammars are written much like in CCG papers. The output is in papers format (well, almost). It is written in Common Lisp. There is an installer in the repo below to get everything needed to set it up.

It implements all established combinators of CCG, also, experimental ones, in all variants and powers. Unary rules, slash modalities, and meta-categories such as (X\X)/X have been implemented. The manual below describes the range of possible CCGlab grammars.

It also implements probabilistic CCG, aka. pCCG. In the plain vanilla version, only lexical features are used, which you can amend. Parameter estimation and update are provided with or without extrapolation. Therefore model training is possible.

In `parsing' mode, CCGlab delivers all logical forms that are derivable for an expression.

In `ranking' mode, for an expression, it gives (1) the most likely logical form, given training from earlier data, (2) its most likely derivation, and (3) the most likely derivation for any LF for the expression. Lower-ranking parsers are available too, if demanded.

Both modes are based on the same CKY parse engine and the same CCG grammar format.

The main features of CCGlab are:

• The logical form is associated with the entire category, via self-contained and lisp-free lambda-calculus.
• Structural unification plays no linguistic role; basic categories and their atomic features are matched for term consistency only.
• There are no built-in or universally assumed basic category inventory. Everything projects from a lexicalized grammar.
• Probabilistic CCG has been integrated into the system. It is the basis of training for ranking.
• By default, standard CCG combinators are available all the time, in their powers (linear, quadratic, cubic). They all have on/off switches for experimentation.
• Facilities are provided for type-raising, such as unary rules and morpholexical generalizations. There is a compiler for it.
• Parsing mode is designed for linguistically-motivated grammars, and ranking mode for statistically-trained grammars. Competence and performance grammars have the same shape but different function. Feature counts are always calculated by CKY independent of mode of use.
• Parsing mode is also possible in a statistically-trained grammar, and ranking mode in a linguistically-motivated grammar. The first one is helpful in fine-tuning a model. The second use is helpful in prioritizing lexical assumptions.

CCGlab resources