Skip to content
Switch branches/tags
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Build Status

gftest: Automatic test case generation for GF grammars

gftest is a program for automatically generating systematic test cases for GF grammars. The basic use case is to give gftest a PGF grammar, a concrete language and a function; then gftest generates a representative and minimal set of example sentences for a human to look at.

There are examples of actual generated test cases later in this document, as well as the full list of options to give to gftest.

It is recommended to compile your PGF with the gf flag --optimize-pgf, otherwise this tool can be very slow. For example, gf -make --optimize-pgf

Table of Contents



Compact form

Check out instructions in this Travis file.

Verbose form

You need the library PGF2, which in turn needs the C runtime for GF. Here are instructions how to install:

  1. Install C runtime: go to the directory gf-core/src/runtime/c, see instructions in INSTALL
  2. Install PGF2 in one of the two ways:
  • EITHER Go to the directory gf-core/src/runtime/haskell-bind, do cabal install
  • OR Go to the root directory of gf-core and compile GF with C-runtime system support: cabal install -fc-runtime, see more information here.

If you get problems, try the following:

  • Remember that you only need the library PGF2, nothing else (e.g. the gf executable doesn't need to work with the C runtime). If you tried installing GF with C-runtime system support, try the lighter version (only installing gf-core/src/runtime/haskell-bind). If that didn't work either, comment out all executables from the cabal file in gf-core/src/runtime/haskell-bind and try again.
  • Libraries not found? Find where they are installed, and try the cabal/stack options --extra-lib-dirs or --extra-include-dirs. For example: stack install --extra-include-dirs="/usr/local/include" --extra-lib-dirs="/usr/local/lib"
    • Using Linux? Try setting up LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib" in your terminal customisation file, such as .bash_profile, .bashrc or .profile
  • You don't know where to find your libraries yourself? Type ghc-pkg list for globally installed packages, and ghc-pkg list --user for local. Then you get an output that starts with a directory on your computer. Go to that directory (e.g. /Users/<your name>/.ghc/<version>/package.conf.d on a mac), open the conf file for the library you're looking for (e.g. /<the previous path>/pgf2- and there you find the path you're looking for under library-dirs:.
  • Any other problem? Contact me: you can e.g. open an issue at GitHub, come to the #gf IRC channel, send an email to gf-dev list or personally to me (

Install gftest

Clone this repository git clone In the root directory (gftest), type cabal install or stack install. This creates an executable gftest.

Common use cases

Run gftest --help of gftest -? to get the list of options.

Common flags:
  -g --grammar=FILE        Path to the grammar (PGF) you want to test
  -l --lang="Eng Swe"      Concrete syntax + optional translations
  -f --function=UseN       Test the given function(s)
  -c --category=NP         Test all functions with given goal category
  -t --tree="UseN tree_N"  Test the given tree
  -s --start-cat=Utt       Use the given category as start category
     --show-cats           Show all available categories
     --show-funs           Show all available functions
     --funs-of-arity=2     Show all functions of arity 2
     --show-coercions      Show coercions in the grammar
     --show-contexts=8140  Show contexts for a given concrete type or a range
                           of concrete types (given as FIds)
     --concr-string=the    Show all functions that include given string
  -q --equal-fields        Show fields whose strings are always identical
  -e --empty-fields        Show fields whose strings are always empty
  -u --unused-fields       Show fields that never make it into the top category
  -r --erased-trees        Show trees that are erased
  -o --old-grammar=ITEM    Path to an earlier version of the grammar
     --only-changed-cats   When comparing against an earlier version of a
                           grammar, only test functions in categories that have
                           changed between versions
     --only-lexicon        When comparing against an earlier version of a
                           grammar, only test lexical categories
  -b --treebank=ITEM       Path to a treebank
     --count-trees=3       Number of trees of size <3>
  -d --debug               Show debug output
  -w --write-to-file       Write the results in a file (<GRAMMAR>_<FUN>.org)
  -? --help                Display help message
  -V --version             Print version information

Grammar: -g

Give the PGF grammar as an argument with -g. If the file is not in the same directory, you need to give the full file path.

It is recommended to compile your PGF with the gf flag --optimize-pgf, otherwise this tool can be very slow. For example, gf -make --optimize-pgf

You can give the grammar with or without .pgf.

Without a concrete syntax you can't do much, but you can see the available categories and functions with --show-cats and --show-funs


  • gftest -g Foods --show-funs
  • gftest -g /home/inari/grammars/LangEng.pgf --show-cats

Language: -l

Give a concrete language. It assumes the format AbsNameConcName, and you should only give the ConcName part.

You can give multiple languages, in which case it will create the test cases based on the first, and show translations in the rest.


  • gftest -g Phrasebook -l Swe --show-cats
  • gftest -g Foods -l "Spa Eng" -f Pizza

Function(s) to test: -f

Given a grammar (-g) and a concrete language ( -l), test a function or several functions.


  • gftest -g Lang -l "Dut Eng" -f UseN
  • gftest -g Phrasebook -l Spa -f "ByTransp ByFoot"

You can use the wildcard *, if you want to match multiple functions. Examples:

  • gftest -g Lang -l Eng -f "*hat*"

matches hat_N, hate_V2, that_Quant, that_Subj, whatPl_IP and whatSg_IP.

  • gftest -g Lang -l Eng -f "*hat*u*"

matches that_Quant and that_Subj.

  • gftest -g Lang -l Eng -f "*"

matches all functions in the grammar. (As of March 2018, takes 13 minutes for the English resource grammar, and results in ~40k lines. You may not want to do this for big grammars.)

Start category for context: -s

Give a start category for contexts. Used in conjunction with -f, -c, -t or --count-trees. If not specified, contexts are created for the start category of the grammar.


  • gftest -g Lang -l "Dut Eng" -f UseN -s Adv

This creates a hole of CN in Adv, instead of the default start category.

Category to test: -c

Given a grammar (-g) and a concrete language ( -l), test all functions that return a given category.


  • gftest -g Phrasebook -l Fre -c Modality
  • gftest -g Phrasebook -l Fre -c ByTransport -s Action

Tree to test: -t

Given a grammar (-g) and a concrete language ( -l), test a complete tree.


  • gftest -g Phrasebook -l Dut -t "ByTransp Bus"

You can combine it with any of the other flags, e.g. put it in a different start category:

  • gftest -g Phrasebook -l Dut -t "ByTransp Bus" -s Action

This may be useful for the following case. Say you tested PrepNP, and the default NP it gave you only uses the word car, but you would really want to see it for some other noun—maybe car_N itself is buggy, and you want to be sure that PrepNP works properly. So then you can call the following:

  • gftest -g TestLang -l Eng -t "PrepNP with_Prep (MassNP (UseN beer_N))"

Compare against an old version of the grammar: -o

Give a grammar, a concrete syntax, and an old version of the same grammar as a separate PGF file. The program generates test sentences for all functions (if no other arguments), linearises with both grammars, and outputs those that differ between the versions. It writes the differences into files.


> gftest -g TestLang -l Eng -o TestLangOld
Created file
Testing functions in…
<categories flashing by>
Created file
Created files TestLangEng-(old|new)
  • All concrete categories that have changed. Shows e.g. if you added or removed a parameter or a field.

  • (usually the most relevant file): All trees that have different linearisations in the following format.

    * send_V3

    ** UseCl (TTAnt TPres ASimul) PPos (PredVP (UsePron we_Pron) (ReflVP (Slash3V3 ∅ (UsePron it_Pron))))
    TestLangDut> we sturen onszelf ernaar
    TestLangDut-OLD> we sturen zichzelf ernaar

    ** UseCl (TTAnt TPast ASimul) PPos (PredVP (UsePron we_Pron) (ReflVP (Slash3V3 ∅ (UsePron it_Pron))))
    TestLangDut> we stuurden onszelf ernaar
    TestLangDut-OLD> we stuurden zichzelf ernaar
  • and groups the functions by their concrete categories. Shows difference if you have e.g. added or removed parameters, and that has created new versions of some functions: say you didn't have gender in nouns, but now you have, then all functions taking nouns have suddenly a gendered version. (This is kind of hard to read, don't worry too much if the output doesn't make any sense.)

Additional arguments to -o

The default mode is to test all functions, but you can also give any combination of -s, -f, -c, --treebank/-b, --only-changed-cats and --only-lexicon.

With -s, you can change the start category in which contexts are generated.

With -f and -c, it tests only the specified functions and categories. With -b FILEPATH (-b=--treebank), it tests only the trees in the file.

With --only-changed-cats, it only test functions in those categories that have changed between the two versions.

With --only-lexicon, it only tests functions that don't take arguments. This mostly corresponds to lexical categories (but not always: it will also catch e.g. hungry_VP in Construction).

Note that this is pretty much overkill when you only want to test lexicon. For big grammars, even this takes a long time, so you might want to try easier alternatives, if lexicon is the only thing you need to test.


  • gftest -g TestLang -l Eng -o old/TestLang tests all functions
  • gftest -g TestLang -l Eng -o old/TestLang -s S tests all functions in start category S
  • gftest -g TestLang -l Eng -o old/TestLang --only-changed-cats tests only changed categories. If no categories have changed (and no other arguments specified), tests everything.
  • gftest -g TestLang -l Eng -o old/TestLang -f "AdjCN AdvCN" -c Adv -b trees.txt tests functions, AdjCN and AdvCN; same for all functions that produce an Adv, and all trees in trees.txt.
  • gftest -g TestLang -l Eng -o old/TestLang --only-lexicon tests only functions without arguments (a good proxy for lexical functions).

Information about a particular string: --concr-string

Show all functions that introduce the string given as an argument.


  • gftest -g Lang -l Eng --concr-string it

which gives the answer ==> CleftAdv, CleftNP, DefArt, ImpersCl, it_Pron

(Note that you have the same feature in GF shell, command morpho_analyse/ma.)

Write into a file: -w

Writes the results into a file of format <GRAMMAR>_<FUN or CAT>.org, e.g. Recommended to open it in emacs org-mode, so you get an overview, and you can maybe ignore some trees if you think they are redundant.

  1. When you open the file, you see a list of generated test cases, like this: Instructions how to use org mode
    Place cursor to the left and click tab to open it.

  2. You get a list of contexts for the test case. Keep the cursor where it was if you want to open everything at the same time. Alternatively, scroll down to one of the contexts and press tab there, if you only want to open one. Instructions how to use org mode

  3. Now you can read the linearisations.
    Instructions how to use org mode

If you want to close the test case, just press tab again, keeping the cursor where it's been all the time (line 31 in the pictures).

Less common use cases

The topics here require some more obscure GF-fu. No need to worry if the terms are not familiar to you.

Empty or always identical fields: -e, -q

Information about the fields: always empty, or always equal to each other. Example of empty fields:

> gftest -g Lang -l Dut -e
* Empty fields:
==> Ant: s

==> Pol: s

==> Temp: s

==> Tense: s

==> V: particle, prefix

The categories Ant, Pol, Temp and Tense are as expected empty; there's no string to be added to the sentences, just a parameter that chooses the right forms of the clause.

V having empty fields particle and prefix is in this case just an artefact of a small lexicon: we happen to have no intransitive verbs with a particle or prefix in the core 300-word vocabulary. But a grammarian would know that it's still relevant to keep those fields, because in some bigger application such a verb may show up.

On the other hand, if some other field is always empty, it might be a hint for the grammarian to remove it altogether.

Example of equal fields:

> gftest -g Lang -l Dut -q
* Equal fields:
==> RCl:
s Pres Simul Pos Utr Pl
s Pres Simul Pos Neutr Pl

==> RCl:
s Pres Simul Neg Utr Pl
s Pres Simul Neg Neutr Pl

==> RCl:
s Pres Anter Pos Utr Pl
s Pres Anter Pos Neutr Pl

==> RCl:
s Pres Anter Neg Utr Pl
s Pres Anter Neg Neutr Pl

==> RCl:
s Past Simul Pos Utr Pl
s Past Simul Pos Neutr Pl

Here we can see that in relative clauses, gender does not seem to play any role in plural. This could be a hint for the grammarian to make a leaner parameter type, e.g. param RClAgr = SgAgr <everything incl. gender> | PlAgr <no gender here>.

Unused fields: -u

These fields are not empty, but they are never used in the top category. The top category can be specified by -s, otherwise it is the default start category of the grammar.

Note that if you give a start category from very low, such as Adv, you get a whole lot of categories and fields that naturally have no way of ever making it into an adverb. So this is mostly meaningful to use for the start category.

Erased trees: -r

Show trees that are erased in some function, i.e. a function F : A -> B -> C has arguments A and B, but doesn't use one of them in the resulting tree of type C. This is usually a bug.


> gftest -g Lang -l "Dut Eng" -r

* Erased trees:

** RelCl (ExistNP something_NP) : RCl
- Tree:  AdvS (PrepNP with_Prep (RelNP (UsePron it_Pron) (UseRCl (TTAnt TPres ASimul) PPos (RelCl (ExistNP something_NP))))) (UseCl (TTAnt TPres ASimul) PPos (ExistNP something_NP))
- Lin:   ermee is er iets
- Trans: with it, such that there is something, there is something

** write_V2 : V2
- Tree:  AdvS (PrepNP with_Prep (PPartNP (UsePron it_Pron) write_V2)) (UseCl (TTAnt TPres ASimul) PPos (ExistNP something_NP))
- Lin:   ermee is er iets
- Trans: with it written there is something

In the first result, an argument of type RCl is missing in the tree constructed by RelNP, and in the second result, the argument write_V2 is missing in the tree constructed by PPartNP. In both cases, the English linearisation contains all the arguments, but in the Dutch one they are missing. (This bug is already fixed, just showing it here to demonstrate the feature.)

Detailed information about the grammar

Debug information: -d

When combined with -f, -c or -t, two things happen:

  1. The trees are linearised using tabularLinearize, which shows the inflection table of all forms.
  2. You can see traces of pruning that happens in testing functions: contexts that are common to several concrete categories are put under a separate test case.

When combined with --show-cats, also the concrete categories are shown.


Shows the categories in the grammar. With --debug/-d, shows also concrete categories.


> gftest -g Foods -l Spa --show-cats -d

* Categories in the grammar:
    Compiles to concrete category   0
    Compiles to concrete categories 1—4
    Compiles to concrete categories 5—6
    Compiles to concrete categories 7—8
    Compiles to concrete category   9


Shows the functions in the grammar. (Nothing fancy happens with other flags.)


First I'll explain what coercions are, then why it may be interesting to show them. Let's take a Spanish Foods grammar, and consider the category Quality, e.g. Good and Vegan. Good "bueno/buena/buenos/buenas" goes before the noun it modifies, whereas Vegan "vegano/vegana/…" goes after, so these will become different concrete categories in the PGF: Quality_before and Quality_after. (In reality, they are something like Quality_7 and Quality_8 though.)

Now, this difference is meaningful only when the adjective is modifying the noun: "la buena pizza" vs. "la pizza vegana". But when the adjective is in a predicative position, they both behave the same: "la pizza es buena" and "la pizza es vegana". For this, the grammar creates a coercion: both Quality_before and Quality_after may be treated as Quality_whatever. To save some redundant work, this coercion Quality_whatever appears in the type of predicative function, whereas the modification function has to be split into two different functions, one taking Quality_before and other Quality_after.

Now you know what coercions are, this is how it looks like in the program:

> gftest -g Foods -l Spa --show-coercions
* Coercions in the grammar:

(Just mentally replace 7 with before, 8 with after and 11 with whatever.)


Show contexts for a given concrete category, given as an FId (i.e. Int). The concrete category may be a coercion or a normal category. By combining with -s, you can change the start category of the context.

(You can get a list of all concrete categories by pairing --show-cats with --debug: see --show-cats.)


  • First, find out some concrete categories:
    > gftest -g Foods -l Spa --show-cats -d
        Compiles to concrete categories 7—8
  • Then, list the contexts for some of them, say Quality_7:
    > gftest -g Foods -l Spa --show-contexts 7

    Pred (That (Mod ∅ Wine)) Vegan
    Pred (That Wine) ∅
    Pred (These (Mod ∅ Wine)) Vegan
    Pred (These Wine) ∅
    Pred (That (Mod ∅ Pizza)) Vegan
    Pred (That Pizza) ∅
    Pred (These (Mod ∅ Pizza)) Vegan
    Pred (These Pizza) ∅
  • You can also give it a range of arguments, with starting and ending category in quotes: gftest -g Foods -l Spa --show-contexts "7 10"

  • Check out from --show-coercions how to find coercions, and you can try --show-contexts with them:

    > gftest -g Foods -l Spa --show-contexts 11

    Pred (That Wine) ∅
    Pred (These Wine) ∅
    Pred (That Pizza) ∅
    Pred (These Pizza) ∅


Number of trees up to given size. Gives a number how many trees, and a couple of examples from the highest size. Examples:

> gftest -g TestLang -l Eng --count-trees 10
There are 675312 trees up to size 10, and 624512 of exactly size 10.
For example:
* AdvS today_Adv (UseCl (TTAnt TPres ASimul) PPos (ExistNP (UsePron i_Pron)))
* UseCl (TTAnt TCond AAnter) PNeg (PredVP (SelfNP (UsePron they_Pron)) UseCopula)

This counts the number of trees in the start category. You can also specify a category:

> gftest -g TestLang -l Eng --count-trees 4 -s Adv
There are 2409 trees up to size 4, and 2163 of exactly size 4.
For example:
* AdAdv very_AdA (PositAdvAdj young_A)
* PrepNP above_Prep (UsePron they_Pron)


Show all functions of given arity (not up to).


> gftest -g Phrasebook --funs-of-arity 3
* Functions in the grammar of arity 3:


Automatic test case generation for GF grammars




No packages published