msyds/gf-core
1
0
Fork 0
forked from GitHub/gf-core
Code Pull Requests Activity
Files
ebef7ecb7be4e6d994fe7b5a9c4aaf98bb44e133
gf-core/doc/Resource-HOWTO.html

968 lines
37 KiB
HTML
Raw Blame History

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML>
<HEAD>
<META NAME="generator" CONTENT="http://txt2tags.sf.net">
<TITLE>Resource grammar writing HOWTO</TITLE>
</HEAD><BODY BGCOLOR="white" TEXT="black">
<P ALIGN="center"><CENTER><H1>Resource grammar writing HOWTO</H1>
<FONT SIZE="4">
<I>Author: Aarne Ranta &lt;aarne (at) cs.chalmers.se&gt;</I><BR>
Last update: Mon Sep 22 14:28:01 2008
</FONT></CENTER>
<P></P>
<HR NOSHADE SIZE=1>
<P></P>
<UL>
<LI><A HREF="#toc1">The resource grammar structure</A>
<UL>
<LI><A HREF="#toc2">Library API modules</A>
<LI><A HREF="#toc3">Phrase category modules</A>
<LI><A HREF="#toc4">Infrastructure modules</A>
<LI><A HREF="#toc5">Lexical modules</A>
</UL>
<LI><A HREF="#toc6">Language-dependent syntax modules</A>
<UL>
<LI><A HREF="#toc7">The present-tense fragment</A>
</UL>
<LI><A HREF="#toc8">Phases of the work</A>
<UL>
<LI><A HREF="#toc9">Putting up a directory</A>
<LI><A HREF="#toc10">Direction of work</A>
<LI><A HREF="#toc11">The develop-test cycle</A>
<LI><A HREF="#toc12">Auxiliary modules</A>
<LI><A HREF="#toc13">Morphology and lexicon</A>
<LI><A HREF="#toc14">Lock fields</A>
<LI><A HREF="#toc15">Lexicon construction</A>
</UL>
<LI><A HREF="#toc16">Lexicon extension</A>
<UL>
<LI><A HREF="#toc17">The irregularity lexicon</A>
<LI><A HREF="#toc18">Lexicon extraction from a word list</A>
<LI><A HREF="#toc19">Lexicon extraction from raw text data</A>
<LI><A HREF="#toc20">Bootstrapping with smart paradigms</A>
</UL>
<LI><A HREF="#toc21">Extending the resource grammar API</A>
<LI><A HREF="#toc22">Using parametrized modules</A>
<UL>
<LI><A HREF="#toc23">Writing an instance of parametrized resource grammar implementation</A>
<LI><A HREF="#toc24">Parametrizing a resource grammar implementation</A>
</UL>
<LI><A HREF="#toc25">Character encoding and transliterations</A>
<LI><A HREF="#toc26">Coding conventions in GF</A>
<LI><A HREF="#toc27">Transliterations</A>
</UL>
<P></P>
<HR NOSHADE SIZE=1>
<P></P>
<P>
<B>History</B>
</P>
<P>
September 2008: updated for Version 1.5.
</P>
<P>
October 2007: updated for Version 1.2.
</P>
<P>
January 2006: first version.
</P>
<P>
The purpose of this document is to tell how to implement the GF
resource grammar API for a new language. We will <I>not</I> cover how
to use the resource grammar, nor how to change the API. But we
will give some hints how to extend the API.
</P>
<P>
A manual for using the resource grammar is found in
</P>
<P>
<A HREF="../lib/resource/doc/synopsis.html"><CODE>www.cs.chalmers.se/Cs/Research/Language-technology/GF/lib/resource/doc/synopsis.html</CODE></A>.
</P>
<P>
A tutorial on GF, also introducing the idea of resource grammars, is found in
</P>
<P>
<A HREF="./gf-tutorial.html"><CODE>www.cs.chalmers.se/Cs/Research/Language-technology/GF/doc/gf-tutorial.html</CODE></A>.
</P>
<P>
This document concerns the API v. 1.5, while the current stable release is 1.4.
You can find the code for the stable release in
</P>
<P>
<A HREF="../lib/resource"><CODE>www.cs.chalmers.se/Cs/Research/Language-technology/GF/lib/resource/</CODE></A>
</P>
<P>
and the next release in
</P>
<P>
<A HREF="../next-lib/src"><CODE>www.cs.chalmers.se/Cs/Research/Language-technology/GF/next-lib/src/</CODE></A>
</P>
<P>
It is recommended to build new grammars to match the next release.
</P>
<A NAME="toc1"></A>
<H2>The resource grammar structure</H2>
<P>
The library is divided into a bunch of modules, whose dependencies
are given in the following figure.
</P>
<P>
<IMG ALIGN="left" SRC="Syntax.png" BORDER="0" ALT="">
</P>
<P>
Modules of different kinds are distinguished as follows:
</P>
<UL>
<LI>solid contours: module seen by end users
<LI>dashed contours: internal module
<LI>ellipse: abstract/concrete pair of modules
<LI>rectangle: resource or instance
<LI>diamond: interface
</UL>
<P>
Put in another way:
</P>
<UL>
<LI>solid rectangles and diamonds: user-accessible library API
<LI>solid ellipses: user-accessible top-level grammar for parsing and linearization
<LI>dashed contours: not visible to users
</UL>
<P>
The dashed ellipses form the main parts of the implementation, on which the resource
grammar programmer has to work with. She also has to work on the <CODE>Paradigms</CODE>
module. The rest of the modules can be produced mechanically from corresponding
modules for other languages, by just changing the language codes appearing in
their module headers.
</P>
<P>
The module structure is rather flat: most modules are direct
parents of <CODE>Grammar</CODE>. The idea
is that the implementors can concentrate on one linguistic aspect at a time, or
also distribute the work among several authors. The module <CODE>Cat</CODE>
defines the "glue" that ties the aspects together - a type system
to which all the other modules conform, so that e.g. <CODE>NP</CODE> means
the same thing in those modules that use <CODE>NP</CODE>s and those that
constructs them.
</P>
<A NAME="toc2"></A>
<H3>Library API modules</H3>
<P>
For the user of the library, these modules are the most important ones.
In a typical application, it is enough to open <CODE>Paradigms</CODE> and <CODE>Syntax</CODE>.
The module <CODE>Try</CODE> combines these two, making it possible to experiment
with combinations of syntactic and lexical constructors by using the
<CODE>cc</CODE> command in the GF shell. Here are short explanations of each API module:
</P>
<UL>
<LI><CODE>Try</CODE>: the whole resource library for a language (<CODE>Paradigms</CODE>, <CODE>Syntax</CODE>,
<CODE>Irreg</CODE>, and <CODE>Extra</CODE>);
produced mechanically as a collection of modules
<LI><CODE>Syntax</CODE>: language-independent categories, syntax functions, and structural words;
produced mechanically as a collection of modules
<LI><CODE>Constructors</CODE>: language-independent syntax functions and structural words;
produced mechanically via functor instantiation
<LI><CODE>Paradigms</CODE>: language-dependent morphological paradigms
</UL>
<A NAME="toc3"></A>
<H3>Phrase category modules</H3>
<P>
The immediate parents of <CODE>Grammar</CODE> will be called <B>phrase category modules</B>,
since each of them concentrates on a particular phrase category (nouns, verbs,
adjectives, sentences,...). A phrase category module tells
<I>how to construct phrases in that category</I>. You will find out that
all functions in any of these modules have the same value type (or maybe
one of a small number of different types). Thus we have
</P>
<UL>
<LI><CODE>Noun</CODE>: construction of nouns and noun phrases
<LI><CODE>Adjective</CODE>: construction of adjectival phrases
<LI><CODE>Verb</CODE>: construction of verb phrases
<LI><CODE>Adverb</CODE>: construction of adverbial phrases
<LI><CODE>Numeral</CODE>: construction of cardinal and ordinal numerals
<LI><CODE>Sentence</CODE>: construction of sentences and imperatives
<LI><CODE>Question</CODE>: construction of questions
<LI><CODE>Relative</CODE>: construction of relative clauses
<LI><CODE>Conjunction</CODE>: coordination of phrases
<LI><CODE>Phrase</CODE>: construction of the major units of text and speech
<LI><CODE>Text</CODE>: construction of texts as sequences of phrases
<LI><CODE>Idiom</CODE>: idiomatic expressions such as existentials
</UL>
<A NAME="toc4"></A>
<H3>Infrastructure modules</H3>
<P>
Expressions of each phrase category are constructed in the corresponding
phrase category module. But their <I>use</I> takes mostly place in other modules.
For instance, noun phrases, which are constructed in <CODE>Noun</CODE>, are
used as arguments of functions of almost all other phrase category modules.
How can we build all these modules independently of each other?
</P>
<P>
As usual in typeful programming, the <I>only</I> thing you need to know
about an object you use is its type. When writing a linearization rule
for a GF abstract syntax function, the only thing you need to know is
the linearization types of its value and argument categories. To achieve
the division of the resource grammar to several parallel phrase category modules,
what we need is an underlying definition of the linearization types. This
definition is given as the implementation of
</P>
<UL>
<LI><CODE>Cat</CODE>: syntactic categories of the resource grammar
</UL>
<P>
Any resource grammar implementation has first to agree on how to implement
<CODE>Cat</CODE>. Luckily enough, even this can be done incrementally: you
can skip the <CODE>lincat</CODE> definition of a category and use the default
<CODE>{s : Str}</CODE> until you need to change it to something else. In
English, for instance, many categories do have this linearization type.
</P>
<A NAME="toc5"></A>
<H3>Lexical modules</H3>
<P>
What is lexical and what is syntactic is not as clearcut in GF as in
some other grammar formalisms. Logically, lexical means atom, i.e. a
<CODE>fun</CODE> with no arguments. Linguistically, one may add to this
that the <CODE>lin</CODE> consists of only one token (or of a table whose values
are single tokens). Even in the restricted lexicon included in the resource
API, the latter rule is sometimes violated in some languages. For instance,
<CODE>Structural.both7and_DConj</CODE> is an atom, but its linearization is
two words e.g. <I>both - and</I>.
</P>
<P>
Another characterization of lexical is that lexical units can be added
almost <I>ad libitum</I>, and they cannot be defined in terms of already
given rules. The lexical modules of the resource API are thus more like
samples than complete lists. There are two such modules:
</P>
<UL>
<LI><CODE>Structural</CODE>: structural words (determiners, conjunctions,...)
<LI><CODE>Lexicon</CODE>: basic everyday content words (nouns, verbs,...)
</UL>
<P>
The module <CODE>Structural</CODE> aims for completeness, and is likely to
be extended in future releases of the resource. The module <CODE>Lexicon</CODE>
gives a "random" list of words, which enables testing the syntax.
It also provides a check list for morphology, since those words are likely to include
most morphological patterns of the language.
</P>
<P>
In the case of <CODE>Lexicon</CODE> it may come out clearer than anywhere else
in the API that it is impossible to give exact translation equivalents in
different languages on the level of a resource grammar. This is no problem,
since application grammars can use the resource in different ways for
different languages.
</P>
<A NAME="toc6"></A>
<H2>Language-dependent syntax modules</H2>
<P>
In addition to the common API, there is room for language-dependent extensions
of the resource. The top level of each languages looks as follows (with German
as example):
</P>
<PRE>
abstract AllGerAbs = Lang, ExtraGerAbs, IrregGerAbs
</PRE>
<P>
where <CODE>ExtraGerAbs</CODE> is a collection of syntactic structures specific to German,
and <CODE>IrregGerAbs</CODE> is a dictionary of irregular words of German
(at the moment, just verbs). Each of these language-specific grammars has
the potential to grow into a full-scale grammar of the language. These grammar
can also be used as libraries, but the possibility of using functors is lost.
</P>
<P>
To give a better overview of language-specific structures,
modules like <CODE>ExtraGerAbs</CODE>
are built from a language-independent module <CODE>ExtraAbs</CODE>
by restricted inheritance:
</P>
<PRE>
abstract ExtraGerAbs = Extra [f,g,...]
</PRE>
<P>
Thus any category and function in <CODE>Extra</CODE> may be shared by a subset of all
languages. One can see this set-up as a matrix, which tells
what <CODE>Extra</CODE> structures
are implemented in what languages. For the common API in <CODE>Grammar</CODE>, the matrix
is filled with 1's (everything is implemented in every language).
</P>
<P>
In a minimal resource grammar implementation, the language-dependent
extensions are just empty modules, but it is good to provide them for
the sake of uniformity.
</P>
<A NAME="toc7"></A>
<H3>The present-tense fragment</H3>
<P>
Some lines in the resource library are suffixed with the comment
</P>
<PRE>
--# notpresent
</PRE>
<P>
which is used by a preprocessor to exclude those lines from
a reduced version of the full resource. This present-tense-only
version is useful for applications in most technical text, since
they reduce the grammar size and compilation time. It can also
be useful to exclude those lines in a first version of resource
implementation. To compile a grammar with present-tense-only, use
</P>
<PRE>
make Present
</PRE>
<P>
with <CODE>resource/Makefile</CODE>.
</P>
<A NAME="toc8"></A>
<H2>Phases of the work</H2>
<A NAME="toc9"></A>
<H3>Putting up a directory</H3>
<P>
Unless you are writing an instance of a parametrized implementation
(Romance or Scandinavian), which will be covered later, the
simplest way is to follow roughly the following procedure. Assume you
are building a grammar for the German language. Here are the first steps,
which we actually followed ourselves when building the German implementation
of resource v. 1.0 at Ubuntu linux. We have slightly modified them to
match resource v. 1.5 and GF v. 3.0.
</P>
<OL>
<LI>Create a sister directory for <CODE>GF/lib/resource/english</CODE>, named
<CODE>german</CODE>.
<PRE>
cd GF/lib/resource/
mkdir german
cd german
</PRE>
<P></P>
<LI>Check out the [ISO 639 3-letter language code
<A HREF="http://www.w3.org/WAI/ER/IG/ert/iso639.htm">http://www.w3.org/WAI/ER/IG/ert/iso639.htm</A>]
for German: both <CODE>Ger</CODE> and <CODE>Deu</CODE> are given, and we pick <CODE>Ger</CODE>.
(We use the 3-letter codes rather than the more common 2-letter codes,
since they will suffice for many more languages!)
<P></P>
<LI>Copy the <CODE>*Eng.gf</CODE> files from <CODE>english</CODE> <CODE>german</CODE>,
and rename them:
<PRE>
cp ../english/*Eng.gf .
rename 's/Eng/Ger/' *Eng.gf
</PRE>
If you don't have the <CODE>rename</CODE> command, you can use a bash script with <CODE>mv</CODE>.
</OL>
<OL>
<LI>Change the <CODE>Eng</CODE> module references to <CODE>Ger</CODE> references
in all files:
<PRE>
sed -i 's/English/German/g' *Ger.gf
sed -i 's/Eng/Ger/g' *Ger.gf
</PRE>
The first line prevents changing the word <CODE>English</CODE>, which appears
here and there in comments, to <CODE>Gerlish</CODE>. The <CODE>sed</CODE> command syntax
may vary depending on your operating system.
<P></P>
<LI>This may of course change unwanted occurrences of the
string <CODE>Eng</CODE> - verify this by
<PRE>
grep Ger *.gf
</PRE>
But you will have to make lots of manual changes in all files anyway!
<P></P>
<LI>Comment out the contents of these files:
<PRE>
sed -i 's/^/--/' *Ger.gf
</PRE>
This will give you a set of templates out of which the grammar
will grow as you uncomment and modify the files rule by rule.
<P></P>
<LI>In all <CODE>.gf</CODE> files, uncomment the module headers and brackets,
leaving the module bodies commented. Unfortunately, there is no
simple way to do this automatically (or to avoid commenting these
lines in the previous step) - but uncommenting the first
and the last lines will actually do the job for many of the files.
<P></P>
<LI>Uncomment the contents of the main grammar file:
<PRE>
sed -i 's/^--//' LangGer.gf
</PRE>
<P></P>
<LI>Now you can open the grammar <CODE>LangGer</CODE> in GF:
<PRE>
gf LangGer.gf
</PRE>
You will get lots of warnings on missing rules, but the grammar will compile.
<P></P>
<LI>At all the following steps you will now have a valid, but incomplete
GF grammar. The GF command
<PRE>
pg -missing
</PRE>
tells you what exactly is missing.
</OL>
<P>
Here is the module structure of <CODE>LangGer</CODE>. It has been simplified by leaving out
the majority of the phrase category modules. Each of them has the same dependencies
as <CODE>VerbGer</CODE>, whose complete dependencies are shown as an example.
</P>
<P>
<IMG ALIGN="middle" SRC="German.png" BORDER="0" ALT="">
</P>
<A NAME="toc10"></A>
<H3>Direction of work</H3>
<P>
The real work starts now. There are many ways to proceed, the most obvious ones being
</P>
<UL>
<LI>Top-down: start from the module <CODE>Phrase</CODE> and go down to <CODE>Sentence</CODE>, then
<CODE>Verb</CODE>, <CODE>Noun</CODE>, and in the end <CODE>Lexicon</CODE>. In this way, you are all the time
building complete phrases, and add them with more content as you proceed.
<B>This approach is not recommended</B>. It is impossible to test the rules if
you have no words to apply the constructions to.
<P></P>
<LI>Bottom-up: set as your first goal to implement <CODE>Lexicon</CODE>. To this end, you
need to write <CODE>ParadigmsGer</CODE>, which in turn needs parts of
<CODE>MorphoGer</CODE> and <CODE>ResGer</CODE>.
<B>This approach is not recommended</B>. You can get stuck to details of
morphology such as irregular words, and you don't have enough grasp about
the type system to decide what forms to cover in morphology.
</UL>
<P>
The practical working direction is thus a saw-like motion between the morphological
and top-level modules. Here is a possible course of the work that gives enough
test data and enough general view at any point:
</P>
<OL>
<LI>Define <CODE>Cat.N</CODE> and the required parameter types in <CODE>ResGer</CODE>. As we define
<PRE>
lincat N = {s : Number =&gt; Case =&gt; Str ; g : Gender} ;
</PRE>
we need the parameter types <CODE>Number</CODE>, <CODE>Case</CODE>, and <CODE>Gender</CODE>. The definition
of <CODE>Number</CODE> in <A HREF="../lib/resource/common/ParamX.gf"><CODE>common/ParamX</CODE></A>
works for German, so we
use it and just define <CODE>Case</CODE> and <CODE>Gender</CODE> in <CODE>ResGer</CODE>.
<P></P>
<LI>Define some cases of <CODE>mkN</CODE> in <CODE>ParadigmsGer</CODE>. In this way you can
already implement a huge amount of nouns correctly in <CODE>LexiconGer</CODE>. Actually
just adding the worst-case instance of <CODE>mkN</CODE> (the one taking the most
arguments) should suffice for every noun - but,
since it is tedious to use, you
might proceed to the next step before returning to morphology and defining the
real work horse, <CODE>mkN</CODE> taking two forms and a gender.
<P></P>
<LI>While doing this, you may want to test the resource independently. Do this by
starting the GF shell in the <CODE>resource</CODE> directory, by the commands
<PRE>
&gt; i -retain german/ParadigmsGer
&gt; cc -table mkN "Kirche"
</PRE>
<P></P>
<LI>Proceed to determiners and pronouns in
<CODE>NounGer</CODE> (<CODE>DetCN UsePron DetQuant NumSg DefArt IndefArt UseN</CODE>) and
<CODE>StructuralGer</CODE> (<CODE>i_Pron this_Quant</CODE>). You also need some categories and
parameter types. At this point, it is maybe not possible to find out the final
linearization types of <CODE>CN</CODE>, <CODE>NP</CODE>, <CODE>Det</CODE>, and <CODE>Quant</CODE>, but at least you should
be able to correctly inflect noun phrases such as <I>every airplane</I>:
<PRE>
&gt; i german/LangGer.gf
&gt; l -table DetCN every_Det (UseN airplane_N)
Nom: jeder Flugzeug
Acc: jeden Flugzeug
Dat: jedem Flugzeug
Gen: jedes Flugzeugs
</PRE>
<P></P>
<LI>Proceed to verbs: define <CODE>CatGer.V</CODE>, <CODE>ResGer.VForm</CODE>, and
<CODE>ParadigmsGer.mkV</CODE>. You may choose to exclude <CODE>notpresent</CODE>
cases at this point. But anyway, you will be able to inflect a good
number of verbs in <CODE>Lexicon</CODE>, such as
<CODE>live_V</CODE> (<CODE>mkV "leben"</CODE>).
<P></P>
<LI>Now you can soon form your first sentences: define <CODE>VP</CODE> and
<CODE>Cl</CODE> in <CODE>CatGer</CODE>, <CODE>VerbGer.UseV</CODE>, and <CODE>SentenceGer.PredVP</CODE>.
Even if you have excluded the tenses, you will be able to produce
<PRE>
&gt; i -preproc=./mkPresent german/LangGer.gf
&gt; l -table PredVP (UsePron i_Pron) (UseV live_V)
Pres Simul Pos Main: ich lebe
Pres Simul Pos Inv: lebe ich
Pres Simul Pos Sub: ich lebe
Pres Simul Neg Main: ich lebe nicht
Pres Simul Neg Inv: lebe ich nicht
Pres Simul Neg Sub: ich nicht lebe
</PRE>
You should also be able to parse:
<PRE>
&gt; p -cat=Cl "ich lebe"
PredVP (UsePron i_Pron) (UseV live_V)
</PRE>
<P></P>
<LI>Transitive verbs
(<CODE>CatGer.V2 CatGer.VPSlash ParadigmsGer.mkV2 VerbGer.ComplSlash VerbGer.SlashV2a</CODE>)
are a natural next step, so that you can
produce <CODE>ich liebe dich</CODE> ("I love you").
<P></P>
<LI>Adjectives (<CODE>CatGer.A ParadigmsGer.mkA NounGer.AdjCN AdjectiveGer.PositA</CODE>)
will force you to think about strong and weak declensions, so that you can
correctly inflect <I>mein neuer Wagen, dieser neue Wagen</I>
("my new car, this new car").
<P></P>
<LI>Once you have implemented the set
(``Noun.DetCN Noun.AdjCN Verb.UseV Verb.ComplSlash Verb.SlashV2a Sentence.PredVP),
you have overcome most of difficulties. You know roughly what parameters
and dependences there are in your language, and you can now proceed very
much in the order you please.
</OL>
<A NAME="toc11"></A>
<H3>The develop-test cycle</H3>
<P>
The following develop-test cycle will
be applied most of the time, both in the first steps described above
and in later steps where you are more on your own.
</P>
<OL>
<LI>Select a phrase category module, e.g. <CODE>NounGer</CODE>, and uncomment some
linearization rules (for instance, <CODE>DetCN</CODE>, as above).
<P></P>
<LI>Write down some German examples of this rule, for instance translations
of "the dog", "the house", "the big house", etc. Write these in all their
different forms (two numbers and four cases).
<P></P>
<LI>Think about the categories involved (<CODE>CN, NP, N, Det</CODE>) and the
variations they have. Encode this in the lincats of <CODE>CatGer</CODE>.
You may have to define some new parameter types in <CODE>ResGer</CODE>.
<P></P>
<LI>To be able to test the construction,
define some words you need to instantiate it
in <CODE>LexiconGer</CODE>. You will also need some regular inflection patterns
in<CODE>ParadigmsGer</CODE>.
<P></P>
<LI>Test by parsing, linearization,
and random generation. In particular, linearization to a table should
be used so that you see all forms produced; the <CODE>treebank</CODE> option
preserves the tree
<PRE>
&gt; gr -cat=NP -number=20 | l -table -treebank
</PRE>
<P></P>
<LI>Save some tree-linearization pairs for later regression testing. You can save
a gold standard treebank and use the Unix <CODE>diff</CODE> command to compare later
linearizations produced from the same list of trees. If you save the trees
in a file <CODE>trees</CODE>, you can do as follows:
<PRE>
&gt; rf -file=trees -tree -lines | l -table -treebank | wf -file=treebank
</PRE>
<P></P>
<LI>A file with trees testing all resource functions is included in the resource,
entitled <CODE>resource/exx-resource.gft</CODE>. A treebank can be created from this by
the Unix command
<PRE>
% runghc Make.hs test langs=Ger
</PRE>
</OL>
<P>
You are likely to run this cycle a few times for each linearization rule
you implement, and some hundreds of times altogether. There are roughly
70 <CODE>cat</CODE>s and
600 <CODE>funs</CODE> in <CODE>Lang</CODE> at the moment; 170 of the <CODE>funs</CODE> are outside the two
lexicon modules).
</P>
<A NAME="toc12"></A>
<H3>Auxiliary modules</H3>
<P>
These auxuliary <CODE>resource</CODE> modules will be written by you.
</P>
<UL>
<LI><CODE>ResGer</CODE>: parameter types and auxiliary operations
(a resource for the resource grammar!)
<LI><CODE>ParadigmsGer</CODE>: complete inflection engine and most important regular paradigms
<LI><CODE>MorphoGer</CODE>: auxiliaries for <CODE>ParadigmsGer</CODE> and <CODE>StructuralGer</CODE>. This need
not be separate from <CODE>ResGer</CODE>.
</UL>
<P>
These modules are language-independent and provided by the existing resource
package.
</P>
<UL>
<LI><CODE>ParamX</CODE>: parameter types used in many languages
<LI><CODE>CommonX</CODE>: implementation of language-uniform categories
such as $Text$ and $Phr$, as well as of
the logical tense, anteriority, and polarity parameters
<LI><CODE>Coordination</CODE>: operations to deal with lists and coordination
<LI><CODE>Prelude</CODE>: general-purpose operations on strings, records,
truth values, etc.
<LI><CODE>Predef</CODE>: general-purpose operations with hard-coded definitions
</UL>
<P>
An important decision is what rules to implement in terms of operations in
<CODE>ResGer</CODE>. The <B>golden rule of functional programming</B> says:
</P>
<UL>
<LI><I>Whenever you find yourself programming by copy and paste, write a function instead!</I>.
</UL>
<P>
This rule suggests that an operation should be created if it is to be
used at least twice. At the same time, a sound principle of <B>vicinity</B> says:
</P>
<UL>
<LI><I>It should not require too much browsing to understand what a piece of code does.</I>
</UL>
<P>
From these two principles, we have derived the following practice:
</P>
<UL>
<LI>If an operation is needed <I>in two different modules</I>,
it should be created in as an <CODE>oper</CODE> in <CODE>ResGer</CODE>. An example is <CODE>mkClause</CODE>,
used in <CODE>Sentence</CODE>, <CODE>Question</CODE>, and <CODE>Relative</CODE>-
<LI>If an operation is needed <I>twice in the same module</I>, but never
outside, it should be created in the same module. Many examples are
found in <CODE>Numerals</CODE>.
<LI>If an operation is needed <I>twice in the same judgement</I>, but never
outside, it should be created by a <CODE>let</CODE> definition.
<LI>If an operation is only needed once, it should not be created as an <CODE>oper</CODE>,
but rather inlined. However, a <CODE>let</CODE> definition may well be in place just
to make the readable.
Most functions in phrase category modules
are implemented in this way.
</UL>
<P>
This discipline is very different from the one followed in early
versions of the library (up to 0.9). We then valued the principle of
abstraction more than vicinity, creating layers of abstraction for
almost everything. This led in practice to the duplication of almost
all code on the <CODE>lin</CODE> and <CODE>oper</CODE> levels, and made the code
hard to understand and maintain.
</P>
<A NAME="toc13"></A>
<H3>Morphology and lexicon</H3>
<P>
The paradigms needed to implement
<CODE>LexiconGer</CODE> are defined in
<CODE>ParadigmsGer</CODE>.
This module provides high-level ways to define the linearization of
lexical items, of categories <CODE>N, A, V</CODE> and their complement-taking
variants.
</P>
<P>
For ease of use, the <CODE>Paradigms</CODE> modules follow a certain
naming convention. Thus they for each lexical category, such as <CODE>N</CODE>,
the overloaded functions, such as <CODE>mkN</CODE>, with the following cases:
</P>
<UL>
<LI>the worst-case construction of <CODE>N</CODE>. Its type signature
has the form
<PRE>
mkN : Str -&gt; ... -&gt; Str -&gt; P -&gt; ... -&gt; Q -&gt; N
</PRE>
with as many string and parameter arguments as can ever be needed to
construct an <CODE>N</CODE>.
<LI>the most regular cases, with just one string argument:
<PRE>
mkN : Str -&gt; N
</PRE>
<LI>A language-dependent (small) set of functions to handle mild irregularities
and common exceptions.
</UL>
<P>
For the complement-taking variants, such as <CODE>V2</CODE>, we provide
</P>
<UL>
<LI>a case that takes a <CODE>V</CODE> and all necessary arguments, such
as case and preposition:
<PRE>
mkV2 : V -&gt; Case -&gt; Str -&gt; V2 ;
</PRE>
<LI>a case that takes a <CODE>Str</CODE> and produces a transitive verb with the direct
object case:
<PRE>
mkV2 : Str -&gt; V2 ;
</PRE>
<LI>A language-dependent (small) set of functions to handle common special cases,
such as transitive verbs that are not regular:
<PRE>
mkV2 : V -&gt; V2 ;
</PRE>
</UL>
<P>
The golden rule for the design of paradigms is that
</P>
<UL>
<LI><I>The user of the library will only need function applications with constants and strings, never any records or tables.</I>
</UL>
<P>
The discipline of data abstraction moreover requires that the user of the resource
is not given access to parameter constructors, but only to constants that denote
them. This gives the resource grammarian the freedom to change the underlying
data representation if needed. It means that the <CODE>ParadigmsGer</CODE> module has
to define constants for those parameter types and constructors that
the application grammarian may need to use, e.g.
</P>
<PRE>
oper
Case : Type ;
nominative, accusative, genitive, dative : Case ;
</PRE>
<P>
These constants are defined in terms of parameter types and constructors
in <CODE>ResGer</CODE> and <CODE>MorphoGer</CODE>, which modules are not
visible to the application grammarian.
</P>
<A NAME="toc14"></A>
<H3>Lock fields</H3>
<P>
An important difference between <CODE>MorphoGer</CODE> and
<CODE>ParadigmsGer</CODE> is that the former uses "raw" record types
for word classes, whereas the latter used category symbols defined in
<CODE>CatGer</CODE>. When these category symbols are used to denote
record types in a resource modules, such as <CODE>ParadigmsGer</CODE>,
a <B>lock field</B> is added to the record, so that categories
with the same implementation are not confused with each other.
(This is inspired by the <CODE>newtype</CODE> discipline in Haskell.)
For instance, the lincats of adverbs and conjunctions are the same
in <CODE>CommonX</CODE> (and therefore in <CODE>CatGer</CODE>, which inherits it):
</P>
<PRE>
lincat Adv = {s : Str} ;
lincat Conj = {s : Str} ;
</PRE>
<P>
But when these category symbols are used to denote their linearization
types in resource module, these definitions are translated to
</P>
<PRE>
oper Adv : Type = {s : Str ; lock_Adv : {}} ;
oper Conj : Type = {s : Str} ; lock_Conj : {}} ;
</PRE>
<P>
In this way, the user of a resource grammar cannot confuse adverbs with
conjunctions. In other words, the lock fields force the type checker
to function as grammaticality checker.
</P>
<P>
When the resource grammar is <CODE>open</CODE>ed in an application grammar, the
lock fields are never seen (except possibly in type error messages),
and the application grammarian should never write them herself. If she
has to do this, it is a sign that the resource grammar is incomplete, and
the proper way to proceed is to fix the resource grammar.
</P>
<P>
The resource grammarian has to provide the dummy lock field values
in her hidden definitions of constants in <CODE>Paradigms</CODE>. For instance,
</P>
<PRE>
mkAdv : Str -&gt; Adv ;
-- mkAdv s = {s = s ; lock_Adv = &lt;&gt;} ;
</PRE>
<P></P>
<A NAME="toc15"></A>
<H3>Lexicon construction</H3>
<P>
The lexicon belonging to <CODE>LangGer</CODE> consists of two modules:
</P>
<UL>
<LI><CODE>StructuralGer</CODE>, structural words, built by using both
<CODE>ParadigmsGer</CODE> and <CODE>MorphoGer</CODE>.
<LI><CODE>LexiconGer</CODE>, content words, built by using <CODE>ParadigmsGer</CODE> only.
</UL>
<P>
The reason why <CODE>MorphoGer</CODE> has to be used in <CODE>StructuralGer</CODE>
is that <CODE>ParadigmsGer</CODE> does not contain constructors for closed
word classes such as pronouns and determiners. The reason why we
recommend <CODE>ParadigmsGer</CODE> for building <CODE>LexiconGer</CODE> is that
the coverage of the paradigms gets thereby tested and that the
use of the paradigms in <CODE>LexiconGer</CODE> gives a good set of examples for
those who want to build new lexica.
</P>
<A NAME="toc16"></A>
<H2>Lexicon extension</H2>
<A NAME="toc17"></A>
<H3>The irregularity lexicon</H3>
<P>
It is useful in most languages to provide a separate module of irregular
verbs and other words which are difficult for a lexicographer
to handle. There are usually a limited number of such words - a
few hundred perhaps. Building such a lexicon separately also
makes it less important to cover <I>everything</I> by the
worst-case variants of the paradigms <CODE>mkV</CODE> etc.
</P>
<A NAME="toc18"></A>
<H3>Lexicon extraction from a word list</H3>
<P>
You can often find resources such as lists of
irregular verbs on the internet. For instance, the
Irregular German Verb page
previously found in
<CODE>http://www.iee.et.tu-dresden.de/~wernerr/grammar/verben_dt.html</CODE>
page gives a list of verbs in the
traditional tabular format, which begins as follows:
</P>
<PRE>
backen (du bäckst, er bäckt) backte [buk] gebacken
befehlen (du befiehlst, er befiehlt; befiehl!) befahl (beföhle; befähle) befohlen
beginnen begann (begönne; begänne) begonnen
beißen biß gebissen
</PRE>
<P>
All you have to do is to write a suitable verb paradigm
</P>
<PRE>
irregV : (x1,_,_,_,_,x6 : Str) -&gt; V ;
</PRE>
<P>
and a Perl or Python or Haskell script that transforms
the table to
</P>
<PRE>
backen_V = irregV "backen" "bäckt" "back" "backte" "backte" "gebacken" ;
befehlen_V = irregV "befehlen" "befiehlt" "befiehl" "befahl" "beföhle" "befohlen" ;
</PRE>
<P></P>
<P>
When using ready-made word lists, you should think about
coyright issues. All resource grammar material should
be provided under GNU Lesser General Public License (LGPL).
</P>
<A NAME="toc19"></A>
<H3>Lexicon extraction from raw text data</H3>
<P>
This is a cheap technique to build a lexicon of thousands
of words, if text data is available in digital format.
See the <A HREF="http://www.cs.chalmers.se/~markus/extract/">Extract Homepage</A>
homepage for details.
</P>
<A NAME="toc20"></A>
<H3>Bootstrapping with smart paradigms</H3>
<P>
This is another cheap technique, where you need as input a list of words with
part-of-speech marking. You initialize the lexicon by using the one-argument
<CODE>mkN</CODE> etc paradigms, and add forms to those words that do not come out right.
This procedure is described in the paper
</P>
<P>
A. Ranta.
How predictable is Finnish morphology? An experiment on lexicon construction.
In J. Nivre, M. Dahllöf and B. Megyesi (eds),
<I>Resourceful Language Technology: Festschrift in Honor of Anna Sågvall Hein</I>,
University of Uppsala,
2008.
Available from the <A HREF="http://publications.uu.se/abstract.xsql?dbid=8933">series homepage</A>
</P>
<A NAME="toc21"></A>
<H2>Extending the resource grammar API</H2>
<P>
Sooner or later it will happen that the resource grammar API
does not suffice for all applications. A common reason is
that it does not include idiomatic expressions in a given language.
The solution then is in the first place to build language-specific
extension modules, like <CODE>ExtraGer</CODE>.
</P>
<A NAME="toc22"></A>
<H2>Using parametrized modules</H2>
<A NAME="toc23"></A>
<H3>Writing an instance of parametrized resource grammar implementation</H3>
<P>
Above we have looked at how a resource implementation is built by
the copy and paste method (from English to German), that is, formally
speaking, from scratch. A more elegant solution available for
families of languages such as Romance and Scandinavian is to
use parametrized modules. The advantages are
</P>
<UL>
<LI>theoretical: linguistic generalizations and insights
<LI>practical: maintainability improves with fewer components
</UL>
<P>
Here is a set of
<A HREF="http://www.cs.chalmers.se/~aarne/geocal2006.pdf">slides</A>
on the topic.
</P>
<A NAME="toc24"></A>
<H3>Parametrizing a resource grammar implementation</H3>
<P>
This is the most demanding form of resource grammar writing.
We do <I>not</I> recommend the method of parametrizing from the
beginning: it is easier to have one language first implemented
in the conventional way and then add another language of the
same family by aprametrization. This means that the copy and
paste method is still used, but at this time the differences
are put into an <CODE>interface</CODE> module.
</P>
<A NAME="toc25"></A>
<H2>Character encoding and transliterations</H2>
<P>
This section is relevant for languages using a non-ASCII character set.
</P>
<A NAME="toc26"></A>
<H2>Coding conventions in GF</H2>
<P>
From version 3.0, GF follows a simple encoding convention:
</P>
<UL>
<LI>GF source files may follow any encoding, such as isolatin-1 or UTF-8;
the default is isolatin-1, and UTF8 must be indicated by the judgement
<PRE>
flags coding = utf8 ;
</PRE>
in each source module.
<LI>for internal processing, all characters are converted to 16-bit unicode,
as the first step of grammar compilation guided by the <CODE>coding</CODE> flag
<LI>as the last step of compilation, all characters are converted to UTF-8
<LI>thus, GF object files (<CODE>gfo</CODE>) and the Portable Grammar Format (<CODE>pgf</CODE>)
are in UTF-8
</UL>
<P>
Most current resource grammars use isolatin-1 in the source, but this does
not affect their use in parallel with grammars written in other encodings.
In fact, a grammar can be put up from modules using different codings.
</P>
<P>
<B>Warning</B>. While string literals may contain any characters, identifiers
must be isolatin-1 letters (or digits, underscores, or dashes). This has to
do with the restrictions of the lexer tool that is used.
</P>
<A NAME="toc27"></A>
<H2>Transliterations</H2>
<P>
While UTF-8 is well supported by most web browsers, its use in terminals and
text editors may cause disappointment. Many grammarians therefore prefer to
use ASCII transliterations. GF 3.0beta2 provides the following built-in
transliterations:
</P>
<UL>
<LI>Arabic
<LI>Devanagari (Hindi)
<LI>Thai
</UL>
<P>
New transliterations can be defined in the GF source file
<A HREF="../src/GF/Text/Transliterations.hs"><CODE>GF/Text/Transliterations.hs</CODE></A>.
This file also gives instructions on how new ones are added.
</P>
<!-- html code generated by txt2tags 2.4 (http://txt2tags.sf.net) -->
<!-- cmdline: txt2tags -\-toc Resource-HOWTO.txt -->
</BODY></HTML>
View Git Blame Copy Permalink