One important reason is agreement, by which I mean that verbs in English, say, agree for number and in a limited way for person. What does this mean in practice? Well, if you’re writing a context-free grammar to handle sentences like “The lady vanishes”, then you can’t just say:

S → NP VP

because that overgenerates. That would allow “The lady vanish”, “The ladies vanishes”, “I vanishes” and so on, because each of these have the form NP VP. “The lady” is an NP (noun phrase), as is “The ladies” and “I”. The rest of these sentences are all VPs (verb phrases). So our grammar has to also say:

S → NP_3rdsg VP_3rdsg

S → NP_non3rdsg VP_non3rdsg

and the same applies to every rule you have in the grammar. Modern grammatical frameworks use feature structures to look after all of this, and enable you to insist that whatever features, like number (singular, plural, and in Slovene dual) or person (I, you, he/she) words have have to agree, so you can write rules like this:

S → NP VP

and let the lexicon, the collection of the words themselves, handle the details.

• ‘S e tidsear a th’ann Calum.
• ‘S e tidsear a th’annam.

The unwary might translate those as “It is a teacher that is in Calum” and “It is a teacher that is in me”, but really tha + ann means “there is”. annam is a preposition marked for person, which I don’t think I’ve mentioned before. I’ve kind of implemented this, but it does overgenerate like mad. Overgeneration is when your grammar allows sentences that aren’t grammatical.

copula.ccg contains the grammar so far. Here are some highlights:This is my feature set so far. EMPH applies to the emphatic and unemphatic forms of the pronoun. EXIS, which doesn’t quite work yet, marks existential clauses.

feature {
NUM<2>: sg pl;
GEND<2>: masc fem;
PERS<2>: 1st 2nd 3rd;
EMPH<2>: emph unemph;
PROP<2>: proper common;
EXIS<2>: exis+ exis-;
}

I have a rather small word list, with tidsear (teacher) and ball-parlamaid as common nouns, and Calum, Somhairle and Sine (pronounced Sheena) as personal names, and is, s and tha as copula, copula and Bi – a special type for a peculiar verb, respectively:

word is:Cop;
word s:Cop;
word tidsear:N;
word ball-parlamaid:N;
word annam:PP;
word Calum:Name;
word Somhairle:Name;
word Sine:Name;
word th:Bi;
word tha:Bi;
word a:Rel;
word pre:P {
ann: exis+;
air: exis-;
}
word pro1:Pro {
mi: 1st sg unemph;
thu: 2nd sg unemph;
e: 3rd sg unemph masc;
i: 3rd sg unemph fem;
}
word pro2:ProEmph {
mise: 1st sg emph;
thusa: 2nd sg emph;
esan: 3rd sg emph masc;
}

and the actual grammatical machinery to get at least some examples working:

family Bi(V) {
entry: s<1>/pp<2>;
}
family P { entry: pp<2>/np<2>; }
family Pro {entry: np<2>[unemph];}
family ProEmph { entry: np<2>[emph]; }
family N { entry: np<2>[common]; }
family Name { entry: np<2>[proper]; }
family Cop(V) {
entry: (s<1>/np<2>[emph])/np<2>[proper];
entry: (s<1>/np<2>[unemph 3rd masc sg])/np<2>[exis+];
}
family NP { entry: np<2>/np<2>; }
family PP { entry: pp<2>; }
family Rel { entry: (np\np)/s; }

I don’t have a particularly smart account of NPs here. But it’s a start. Also, I can’t work out how to get apostrophes in.

To do: make sure that it doesn’t overgenerate for non-existential clauses. I’m not sure, for example, that tha air Calum is an S, although tha ann Calum is.

• You may not already have a recent version of python, which you can get from http://www.python.org/download/releases/2.7.1/ as a .dmg, which has a friendly hand-holdy installation process.
• Environmental variables:
  • export JAVA_HOME=/usr (this surprised me, but it works on Mac OS X 10.4.11)
  • export PATH=”$PATH:$OPENCCG_HOME/bin”
  • cd to the directory that you’ve downloaded openccg to, type pwd, and set OPENCCG_HOME to it using export.
• You will also need to fetch lex.py and yacc.py from sourceforge: http://openccg.cvs.sourceforge.net/viewvc/openccg/openccg/bin/ and put them in the bin folder in your OpenCCG installation.
• If you then follow the instructions in the README file and get an error about the wrong class number you’ll have to rebuild it. Try typing ant at the command line and see what happens. I don’t remember installing ant, which means that it might come on the Mac by default. If not, you’ll have to go to http://ant.apache.org/. Good luck!

It comes with some minuscule test grammars including Basque and Turkish.

The number of prepositions is unusually high and indicates that PPs (prepositional phrases) do an awful lot of the work in a Gaelic sentence. The number of verbs seems pretty low, and in fact many of them are forms of the verbs “to be” that we’ve seen earlier. This is because the verb “to be” typically does much of the rest of the work. More examples of this to come.

The article doesn’t mark gender (of which there are two, masculine and feminine) but it does mark the two numbers (singular and plural). So how come there are five articles listed?

Well, an is the singular, na does double duty for “of the” and “the” plural. nan does “of the” plural. Before a labial consonant, an becomes am and nan becomes nam. This warns us that our system will have to take into account initial consonants to get this right.

There are also some duplicates. “Scotland” is Alba normally and h-Alba after na, as in Banca na h-Alba “Bank of Scotland”. duine (person) has a weird-looking plural, daoine. dùthaich has the genitive form dùthcha. baile (town) has a lenited form (I will come to this, but not today) bhaile. So we see that Gaelic is not only morphologically rich, but instead of adding case endings and whatnot to the ends of words, like in Hungarian or Turkish, modifies the insides of words instead.

That will do for the now.

Now I can do word-token counts on it, using terrible code like the following:

#!/usr/bin/perl -w
my %list;
while (<>) {
@tokens = split(/[\s\)\(\.=\,\?]/);
for $token (@tokens)
{
$list{lc($token)}++;
}
}
foreach $key (sort { $list{$b} <=> $list{$a} } (keys %list)) {
print " $key $list{$key}\n";
}


Note the entirely ad hoc collection of characters to split on.
The list is here, and you will see that the first noun is baile (town) at number 25, which tells you more about Wikipedia than it does about Gaelic. But also that an, e and is are, as we have seen, ambiguous between parts of speech, and that I can’t quite work out what to do with a.

• Positive: Is mise Calum. (I am Calum.)
• Interrogative: An tusa Ealasaid? (Are you Elizabeth?)
• Negative: Cha mise Calum. (I amn’t Calum.)
• Negative interrogative: Chan esan Uilleam? (Aren’t you William?)

However, is doesn’t have type (S/NP)/NP because you can’t say

*Is Calum tidsear. (Calum is a teacher.)

(The star indicates that a sentence is ungrammatical.) You also can’t say

* Is mise tidsear. (I am a teacher.)

So we can, for now, assign it type (S/Nper)/PN, where Nper is a personal pronoun and PN is a proper noun. But let me assure you it gets much harder.

Coming soon: How to say that I am a teacher.

X/Y Y → X

So I, being simple minded, wrote a bit of code that

• Tokenized a sentence on spaces, and separated out punctuation marks (actually just the question mark).
• Converted that tokenlist into a long string of types.
• Applied forward composition to remove any like types next to each other and the slash that preceded them.
• Printed out what remains.

My sentences are:

• Dè tha dol? What goes? What’s happening?
• An robh fhios agaibh? Did you know?

Let’s try the first one. The output of the program looks like this:

S[int]/?/VPVP/VNVN?
S[int]/?/VPVP?
S[int]/??
S[int]

based on the following types:

• Dè → S[int]/?/VP
  • S[int] indicates that the sentence is interrogative.
  • /? is there to digest the question mark at the end.
• tha → VP/VN
  • give me a verbal noun, of which more soon, and I will give you a VP
• dol → VN
• ? → ?

I’m not happy with this grammar, which I wrote on Friday evening. But I leave it here for demonstration purposes.

Observations:

1. This is a bit of a cheat, because the code doesn’t actually give you a proper bracketed parse.
2. Also, it only allows for each lexical item to have a single type, which is wrong.
3. But, despite all that and it being almost the simplest piece of code I’ve ever written, it identifies that those sentences are well-formed. If you try feeding it a sentence with the same words in a different order, it doesn’t give you S[int] out at the end.
4. Something’s gone horribly wrong with the character encoding. You will just have to imagine the letter è in Dè tha dol? But I’m not going to investigate at this time of night.
5. Code in haste, blog at leisure. It’s taken longer to write these 300-odd words than it did to produce the code in the first place.

Available here: ForwardComposition.java

Coming soon: To be and to be (2).

Bi has three forms in the present tense, according to whether it’s positive, interrogative, negative or negative interrogative, all thanks to particles which I don’t know whether they’re a VMOD or a P or a what.

• Positive: Tha mi sgìth (I am tired). Tha is independent so is (S/NP)/ADJ.
• Negative: Chan eil mi sgìth (I am not tired). Eil is dependent, so its type will be ((S\VMOD)/NP)/ADJ.
• Negative interrogative: Nach eil thu sgìth? (Are you not tired?) Same as above.
• Interrogative: A bheil thu sgìth? (Are you tired?) Bheil is dependent again but whereas most verbs just have a dependent and an independent form in any one tense, bi has two dependent forms. So, what to do?

Whatever happens, we have a tree at the top which says something like (Vspec Vbar). I don’t especially like this because eil thu sgìth, the Vbar, isn’t a constituent. I’m willing to lay money that there are no song titles that begin “Eil”. A coordination test for constituency, incidentally, isn’t decisive in English because one thing that categorial grammar is good at is non-constituent coordination, say in “Mary loves pizza and Tim rice”.

So either we commit ourselves to a type S/Vbar for a or chan or nach, which is potentially good because you could parse the entire sentence with forward composition, of which more tomorrow, or to assign eil type ((S\Vspec_neg)/NP)/ADJ and bheil type ((S\Vspec_int)/NP/ADJ. More types will be needed for all of these forms, because bi isn’t just for adjectives!

Tomorrow: what is the simplest parser that could possibly work?

Aside: hopefully these are the last examples I give in English.

“Mary loves pizza”. “Mary” is a singular personal name, “pizza” is a mass noun, and those are both sorts of NP. What about “loves”? It’s (S\NP)/NP. A simpler example is “Ice melts”. “Ice” is an NP, and “melts” here is S\NP. The backslash in Y\X means “give me something of type X to my left and I’ll give you a Y“.

So (S\NP)/NP, with a forward slash and a backslash, takes NPs to the left and right, and gives you an S, or a sentence.

In principle Gaelic verbs should have type (S/NP)/NP, but I have never seen a sentence exactly like this. “Mary loves pizza”, after all, is only OK because “loves” is stative. Unless you do the marketing for McDonald’s.