Will Price Posts About

DCG object level representation in prolog

Prolog is an excellent language for writing parsers using Definite Clause Grammars (DCG) which are supported as first class objects in the language. The language provides syntactic sugar to make writing these grammars more easily than without.

Writing a DCG isn’t far off translating a given grammar into prolog syntax, but before we get to that we’ll first look at writing DCGs the hard way without syntactic sugar. It’s important to understand the object level representation of a DCG as this is what prolog actually executes.

Parsing a sentence using a DCG makes use of a standard prolog programming pattern: the difference list. The difference list programming pattern represents a list L as the difference between two other lists L1 and L2 where L2 is a sublist of L1 and append(L, L2, L1) (that is, L appended with L2 forms L1).

A list in difference list form is represented by two lists. Let’s say we want to express L = [element1, element2, ..., elementN] in difference list form. We need two lists L1 and L2 in which we’ll encode L as the difference between those lists. We’ll represent the two lists as a pair, pairs in prolog use the syntax A-B where A is the first element of the pair, and B the second. This is not necessary, and later we’ll drop the pair syntax and just use arguments to hold each of the lists. Getting back to encoding L using L1 and L2 we can form the difference list L1-L2 = [element1, element2, ..., elementN | Rest]-Rest. L1 is L prepended to L2, and L2 is some other list, we don’t really care what.

“Why do this, you’re insane!?” You might ask, well we now have a handle Rest on the end of the list which allows us to do some clever tricks to write efficient predicates, and it makes writing parsers super easy.

Appending with difference lists

One of the simplest examples of the difference list pattern is an implementation of the append/3 predicate. You may have seen the naive implementation:

append([], L, L).
append([Head|Tail],List,[Head|Rest]) :- append(Tail, List, Rest).

However this takes linear time in execution… for something as frequently used as appending, this isn’t good enough, we want constant time execution! To achieve constant time execution of appending we’ll implement append/3 using difference lists.

Let’s start with a simple example, appending [1, 2, 3] to [4, 5, 6, 7]. First lets convert these to difference list representation: [1, 2, 3 | Rest1]-Rest1, [4, 5, 6, 7 | Rest2]-Rest2 (suspend your suspicion that this is a pointless, useless representation, we’ll get there!). What should the result of appending the two lists result in (in difference list form?)? Well it should at least start with [1, 2, 3, 4, 5, 6, 7 | Something]-Something as this contains the result, but how about the Rest part of the list which we’ve left as Something? I suppose we could put anything here but there’s an alternative that makes a bit more sense: reuse the Rest part from the second list, this makes a bit more sense, as we preserve more information about our inputs this way. In conclusion we’ve decided that appending the difference lists [1, 2, 3 | Rest1]-Rest1 and [4, 5, 6, 7 | Rest2]-Rest2 should result in [1, 2, 3, 4, 5, 6, 7 | Rest2]-Rest2.

How we write append/3 such that it uses difference lists? Lets distinguish it from the recursive definition we presented above my denoting it as dappend/3 (for difference-list append). We know the inputs, so they form the first 2 arguments of the predicate, and the last argument will be the appended difference lists. Let’s sketch this out:

dappend([L1element1, L1element2, ..., L1elementN | Rest1]-Rest1,
        [L2element1, L2element2, ..., L2elementM | Rest2]-Rest2,
        [L1element1, L1element2, ..., L1elementN, L2element1, L2element2, ..., L2elementM | Rest2]-Rest2]).

We could write out a bunch of predicates for all different sizes of lists, but that’d be insanity! Instead we can make clever use of the Rest parts of the difference lists to combine them!

dappend(L1-Rest1, L2-Rest2, L1-Rest2) :- Rest1 = L2.
% or more simply:
dappend(L1-L2, L2-Rest, L1-Rest).

By unifying Rest1 with L2 we force the difference part of the L1 difference list to be the L2 difference list. Going back to our example, we unify Rest1 in [1, 2, 3 | Rest1]-Rest1 with [4, 5, 6, 7 | Rest2] resulting in [1, 2, 3, 4, 5, 6, 7 | Rest2], then make this into a difference list by lifting it into the pair functor -/2 in the output argument [1, 2, 3, 4, 5, 6, 7 | Rest2]-Rest2. No recursion, just unification!

Now what if we want to wrap this up so we don’t have to use difference lists directly? Let’s try and write a wrapper my_append/3 that takes 2 normal lists L1 and L2 and concatenates them to form L, first we have to convert L1 and L2 into difference lists, then we can call dappend/3 and pull out the difference of the result to get L.

diff_list([], End-End).
diff_list([Head|Rest], [Head|DiffListRest]-End) :-
  diff_list(Rest, DiffListRest-End).

my_append(L1, L2, L) :-
  diff_list(L1, DiffList1),
  dappend(L1-_, L2-_, L-[]).

And some examples using dappend/3 and my_append/3:

?- L2 = [4, 5, 6], dappend([1,2,3|L2]-L2, L2-[], L-[]).
L = [1, 2, 3, 4, 5, 6].
% We force the difference list representing `L1` to be the
% difference between L1 appended to L2.

?- dappend([1, 2, 3]-_, [4, 5, 6]-[], L-[]).
L = [1, 2, 3].
% By failing to create a proper difference list in the L1 position, namely
% `[1, 2, 3]-_`, `_` unifies with `[4, 5, 6]` and L unifies with `[1, 2, 3]`.

To make the normal lists into difference lists we have to traverse them to the end… defeating the whole object of the exercise: to make append/3 work in constant time, so DON’T use this unless you already have your lists in difference list form.

Exercise: Take a few example difference lists and try appending them with the dappend/3 predicate.

DCGs: Grammars with difference lists

You should now be familiar with the concept of difference lists from the dappend/3 example. We now show how to construct parsers from difference lists. What does it mean to parse a sentence? How are we going to represent a sentence? What does it mean to partially parse a sentence? Keep these questions in mind.

Prolog represents sentences as lists of words, an example being [hello, my, name, is, will]. How about parsing this sentence? To parse it, we need a grammar. A grammar specifies the legal sentences in the language; we’ll constrain our language to be sentences of the form hello my name is <name> where <name> is either will, tom, or henry.

Lets define this in EBNF, a standard form for representing grammars.

sentence = greeting, introduction;
greeting = "hello";
introduction = "my", "name", "is", name;
name = "will" | "tom" | "henry";

What are the valid sentences represented in Prolog?

[hello, my, name, is, will]
[hello, my, name, is, tom]
[hello, my, name, is, henry]

Parsing the sentence top-down from sentence involves substituting the LHS of a rule for the RHS repeatedly until all non-terminals are replaced with terminals. [hello, my name, is will] parsed by the rule sentence takes the following sequence:

  • [], sentence
  • [], greeting, introduction
  • [hello], introduction
  • [hello, my, name, is], name
  • [hello, my, name, is, will]

At each stage we replace a non-terminal with the RHS of its rule, we accumulate the non-terminals into a list starting with the empty list. Parsers consume terminals, that is given a sentence, rules will match part of the sentence.

How would we parse the greeting rule in prolog using difference lists? We represent the parsed terminals as the difference between two lists: greeting(L1, L2) the difference between L1 and L2 should be hello, we can write this as greeting([hello|Rest], Rest), we match hello from the head of the first list, consuming it, leaving Rest.

Let’s see how this predicate functions in a few queries:

?- greeting([hello], Rest).
Rest = []

?- greeting([hello, my, name, is ,will], Rest).
Rest = [my, name is will]

?- greeting(Sentence, [my, name, is will]).
Sentence = [hello, my, name, is will]

?- greeting([], [])

?- greeting([hello], [])

?- greeting([hello, my, name, is will], [])

Although greeting was arguably the simplest of the grammar rules, the concept applies to all other grammar rules, all we’re doing is consuming terminals from the first list, putting the resulting list in the second argument.

Let’s convert the rest of the grammar rules to Prolog in difference list form.

name is very similar to greeting, we simply consume the name.

name([will|Rest], Rest).
name([tom|Rest], Rest).
name([henry|Rest], Rest).

introduction is slightly more complicated as now we have a non-terminal in the RHS of the rule

introduction([my, name, is|Rest1], Rest2) :- name(Rest1, Rest2).

This deserves some explanation. introduction consumes the non-terminals on the RHS of its corresponding grammar rule, but then we also need to give the rest of the sentence to the name parser so it can check that the rest of the sentence is a name. To do this we pass the remainder of the sentence, Rest1, onto name which results in another list, containing the remainder of the sentence after matching the name in Rest2. Let’s give some pseudo-Prolog examples to further demonstrate how this works:

% This is not a real prolog session, I've written out the intermediate
% variables to aid the understanding of which parts of the sentence are
% consumed by which parsers

?- introduction([my, name, is, will], X).
Rest1 = [will],
Rest2 = [],
X = Rest2 = []

?- introduction([my], X).
% Prolog can't match the [my] to [my, name, is|Rest1] as the list is missing [name, is]

?- introduction([my, name, is, Name], []).
Name = will;
Name = tom;
Name = henry

Finally the sentence rule:

sentence(L1, L3) :- greeting(L1, L2), introduction(L2, L3).

and some examples:

% This is not a real prolog session, I've written out the intermediate
% variables to aid the understanding of which parts of the sentence are
% consumed by which parsers

?- sentence([hello, my, name, is will], []).
L1 = [hello, my, name, is will],
L2 = [my, name, is will],
L3 = []
% greeting/2 consumes `hello` from the list leaving L2,
% which is then fully consumed by introduction/2 resulting
% in L3 = []

?- sentence(Sentence, []).
Sentence = [hello, my, name, is will];
Sentence = [hello, my, name, is tom];
Sentence = [hello, my, name, is henry].

?- sentence([my, name, is, will], [will]).
% Because name is called like `name([will|Rest], [will])`, it doesn't match the
% definition of `name/2` so fails.

Translating these grammar rules is very mechanical, which is why Prolog introduces syntactic sugar for them! Yay! The grammar we gave before can be rewritten as:

sentence --> greeting, introduction.
greeting --> [hello].
introduction --> [hello, my, name, is], name.
name --> [will].
name --> [tom].
name --> [henry].

We also don’t have to explicitly call the predicates like sentence(Sentence, []), we can use phrase/2 instead: phrase(sentence, Sentence) is equivalent to the previous call.

More resources