Using the Python binding to the C runtime

Krasimir Angelov, July 2015

Loading the Grammar

Before you use the Python binding you need to import the pgf module.
>>> import pgf
Once you have the module imported, you can use the dir and help functions to see what kind of functionality is available. dir takes an object and returns a list of methods available in the object:
>>> dir(pgf)
help is a little bit more advanced and it tries to produce more human readable documentation, which more over contains comments:
>>> help(pgf)
A grammar is loaded by calling the method readPGF:
>>> gr = pgf.readPGF("App12.pgf")
From the grammar you can query the set of available languages. It is accessible through the property languages which is a map from language name to an object of class pgf.Concr which respresents the language. For example the following will extract the English language:
>>> eng = gr.languages["AppEng"]
>>> print(eng)
<pgf.Concr object at 0x7f7dfa4471d0>


All language specific services are available as methods of the class pgf.Concr. For example to invoke the parser, you can call:
>>> i = eng.parse("this is a small theatre")
This gives you an iterator which can enumerates all possible abstract trees. You can get the next tree by calling next:
>>> p,e =
or by calling __next__ if you are using Python 3:
>>> p,e = i.__next__()
The results are always pairs of probability and tree. The probabilities are negated logarithmic probabilities and which means that the lowest number encodes the most probable result. The possible trees are returned in decreasing probability order (i.e. increasing negated logarithm). The first tree should have the smallest p:
>>> print(p)
and this is the corresponding abstract tree:
>>> print(e)
PhrUtt NoPConj (UttS (UseCl (TTAnt TPres ASimul) PPos (PredVP (DetNP (DetQuant this_Quant NumSg)) (UseComp (CompNP (DetCN (DetQuant IndefArt NumSg) (AdjCN (PositA small_A) (UseN theatre_N)))))))) NoVoc
The parse method has also the following optional parameters:
catstart category
nmaximum number of trees
heuristicsa real number from 0 to 1
callbacksa list of category and callback function
By using these parameters it is possible for instance to change the start category for the parser or to limit the number of trees returned from the parser. For example parsing with a different start category can be done as follows:
>>> i = eng.parse("a small theatre", cat="NP")

The heuristics factor can be used to trade parsing speed for quality. By default the list of trees is sorted by probability this corresponds to factor 0.0. When we increase the factor then parsing becomes faster but at the same time the sorting becomes imprecise. The worst factor is 1.0. In any case the parser always returns the same set of trees but in different order. Our experience is that even a factor of about 0.6-0.8 with the translation grammar, still orders the most probable tree on top of the list but further down the list the trees become shuffled.

The callbacks is a list of functions that can be used for recognizing literals. For example we use those for recognizing names and unknown words in the translator.


You can either linearize the result from the parser back to another language, or you can explicitly construct a tree and then linearize it in any language. For example, we can create a new expression like this:
>>> e = pgf.readExpr("AdjCN (PositA red_A) (UseN theatre_N)")
and then we can linearize it:
>>> print(eng.linearize(e))
red theatre
This method produces only a single linearization. If you use variants in the grammar then you might want to see all possible linearizations. For that purpouse you should use linearizeAll:
>>> for s in eng.linearizeAll(e):
red theatre
red theater
If, instead, you need an inflection table with all possible forms then the right method to use is tabularLinearize:
>>> eng.tabularLinearize(e):
{'s Sg Nom': 'red theatre', 's Pl Nom': 'red theatres', 's Pl Gen': "red theatres'", 's Sg Gen': "red theatre's"}

Finally, you could also get a linearization which is bracketed into a list of phrases:

>>> [b] = eng.bracketedLinearize(e)
>>> print(b)
(CN:4 (AP:1 (A:0 red)) (CN:3 (N:2 theatre)))
Each bracket is actually an object of type pgf.Bracket. The property cat of the object gives you the name of the category and the property children gives you a list of nested brackets. If a phrase is discontinuous then it is represented as more than one brackets with the same category name. In that case, the index that you see in the example above will have the same value for all brackets of the same phrase.

The linearization works even if there are functions in the tree that doesn't have linearization definitions. In that case you will just see the name of the function in the generated string. It is sometimes helpful to be able to see whether a function is linearizable or not. This can be done in this way:
>>> print(eng.hasLinearization("apple_N"))

Analysing and Constructing Expressions

An already constructed tree can be analyzed and transformed in the host application. For example you can deconstruct a tree into a function name and a list of arguments:

>>> e.unpack()
('AdjCN', [<pgf.Expr object at 0x7f7df6db78c8>, <pgf.Expr object at 0x7f7df6db7878>])
The result from unpack can be different depending on the form of the tree. If the tree is a function application then you always get a tuple of function name and a list of arguments. If instead the tree is just a literal string then the return value is the actual literal. For example the result from:
>>> pgf.readExpr('"literal"').unpack()
is just the string 'literal'. Situations like this can be detected in Python by checking the type of the result from unpack.

For more complex analyses you can use the visitor pattern. In object oriented languages this is just a clumpsy way to do what is called pattern matching in most functional languages. You need to define a class which has one method for each function in the abstract syntax of the grammar. If the functions is called f then you need a method called on_f. The method will be called each time when the corresponding function is encountered, and its arguments will be the arguments from the original tree. If there is no matching method name then the runtime will to call the method default. The following is an example:

>>> class ExampleVisitor:
		def on_DetCN(self,quant,cn):
			print("Found DetCN")
		def on_AdjCN(self,adj,cn):
			print("Found AdjCN")
		def default(self,e):
>>> e2.visit(ExampleVisitor())
Found DetCN
Found AdjCN
Here we call the method visit from the tree e2 and we give it, as parameter, an instance of class ExampleVisitor. ExampleVisitor has two methods on_DetCN and on_AdjCN which are called when the top function of the current tree is DetCN or AdjCN correspondingly. In this example we just print a message and we call visit recursively to go deeper into the tree.

Constructing new trees is also easy. You can either use readExpr to read trees from strings, or you can construct new trees from existing pieces. This is possible by using the constructor for pgf.Expr:
>>> quant = pgf.readExpr("DetQuant IndefArt NumSg")
>>> e2 = pgf.Expr("DetCN", [quant, e])
>>> print(e2)
DetCN (DetQuant IndefArt NumSg) (AdjCN (PositA red_A) (UseN theatre_N))

Embedded GF Grammars

The GF compiler allows for easy integration of grammars in Haskell applications. For that purpose the compiler generates Haskell code that makes the integration of grammars easier. Since Python is a dynamic language the same can be done at runtime. Once you load the grammar you can call the method embed, which will dynamically create a Python module with one Python function for every function in the abstract syntax of the grammar. After that you can simply import the module:
>>> gr.embed("App")
<module 'App' (built-in)>
>>> import App
Now creating new trees is just a matter of calling ordinary Python functions:
>>> print(App.DetCN(quant,e))
DetCN (DetQuant IndefArt NumSg) (AdjCN (PositA red_A) (UseN house_N))

Access the Morphological Lexicon

There are two methods that gives you direct access to the morphological lexicon. The first makes it possible to dump the full form lexicon. The following code just iterates over the lexicon and prints each word form with its possible analyses:
for entry in eng.fullFormLexicon():
The second one implements a simple lookup. The argument is a word form and the result is a list of analyses:
[('letter_1_N', 's Sg Nom', inf), ('letter_2_N', 's Sg Nom', inf)]

Access the Abstract Syntax

There is a simple API for accessing the abstract syntax. For example, you can get a list of abstract functions:
>>> gr.functions
or a list of categories:
>>> gr.categories
You can also access all functions with the same result category:
>>> gr.functionsByCat("Weekday")
['friday_Weekday', 'monday_Weekday', 'saturday_Weekday', 'sunday_Weekday', 'thursday_Weekday', 'tuesday_Weekday', 'wednesday_Weekday']
The full type of a function can be retrieved as:
>>> print(gr.functionType("DetCN"))
Det -> CN -> NP

Type Checking Abstract Trees

The runtime type checker can do type checking and type inference for simple types. Dependent types are still not fully implemented in the current runtime. The inference is done with method inferExpr:

>>> e,ty = gr.inferExpr(e)
>>> print(e)
AdjCN (PositA red_A) (UseN theatre_N)
>>> print(ty)
The result is a potentially updated expression and its type. In this case we always deal with simple types, which means that the new expression will be always equal to the original expression. However, this wouldn't be true when dependent types are added.

Type checking is also trivial:

>>> e = gr.checkExpr(e,pgf.readType("CN"))
>>> print(e)
AdjCN (PositA red_A) (UseN theatre_N)
In case of type error you will get an exception:
>>> e = gr.checkExpr(e,pgf.readType("A"))
pgf.TypeError: The expected type of the expression AdjCN (PositA red_A) (UseN theatre_N) is A but CN is infered

Partial Grammar Loading

By default the whole grammar is compiled into a single file which consists of an abstract syntax together will all concrete languages. For large grammars with many languages this might be inconvinient because loading becomes slower and the grammar takes more memory. For that purpose you could split the grammar into one file for the abstract syntax and one file for every concrete syntax. This is done by using the option -split-pgf in the compiler:
$ gf -make -split-pgf App12.pgf
Now you can load the grammar as usual but this time only the abstract syntax will be loaded. You can still use the languages property to get the list of languages and the corresponding concrete syntax objects:
>>> gr = pgf.readPGF("App.pgf")
>>> eng = gr.languages["AppEng"]
However, if you now try to use the concrete syntax then you will get an exception:
>>> gr.languages["AppEng"].lookupMorpho("letter")
Traceback (most recent call last):
  File "", line 1, in 
pgf.PGFError: The concrete syntax is not loaded
Before using the concrete syntax, you need to explicitly load it:
>>> eng.load("AppEng.pgf_c")
>>> print(eng.lookupMorpho("letter"))
[('letter_1_N', 's Sg Nom', inf), ('letter_2_N', 's Sg Nom', inf)]
When you don't need the language anymore then you can simply unload it:
>>> eng.unload()


GraphViz is used for visualizing abstract syntax trees and parse trees. In both cases the result is a GraphViz code that can be used for rendering the trees. See the examples bellow.
>>> print(gr.graphvizAbstractTree(e))
graph {
n0[label = "AdjCN", style = "solid", shape = "plaintext"]
n1[label = "PositA", style = "solid", shape = "plaintext"]
n2[label = "red_A", style = "solid", shape = "plaintext"]
n1 -- n2 [style = "solid"]
n0 -- n1 [style = "solid"]
n3[label = "UseN", style = "solid", shape = "plaintext"]
n4[label = "theatre_N", style = "solid", shape = "plaintext"]
n3 -- n4 [style = "solid"]
n0 -- n3 [style = "solid"]
>>> print(eng.graphvizParseTree(e))
graph {

  subgraph {rank=same;

  subgraph {rank=same;
    n1 -- n3
  n4 -- n1
  n4 -- n3

  subgraph {rank=same;
    n0 -- n2
  n1 -- n0
  n3 -- n2

  subgraph {rank=same;
    n100000 -- n100001
  n0 -- n100000
  n2 -- n100001