| Lingua-FeatureMatrix documentation | Contained in the Lingua-FeatureMatrix distribution. |
Lingua::FeatureMatrix - Perl extension for configuring groups of (e.g.) phonemes into feature groups
use Lingua::FeatureMatrix;
# this example uses the module provided in the examples directory of
# the distro; you'll want to create your own 'Eme' subclass or
# modify 'Phone.pm' for yourself:
use lib 'examples';
use Phone;
# construct a new feature-matrix from a dat file (here using dat
# file same as example below)
my $matrix =
Lingua::FeatureMatrix->new(eme => Phone,
file => 'examples/phonematrix.dat');
if ($matrix->matchesFeatureClass('EE', 'VOW')) {
# EE is a "vow", bless this properly
push @Pope::ISA, 'Catholic';
}
if (not $matrix->matchesFeatureClass('AA', 'AFF')) {
# will be executed
$deadman->walking();
}
if ($matrix->matchesFeatureClass('S', 'VOW')) {
# won't happen
map { $_->fly() } @pigs;
}
# silliness aside, you can also dump a filled-out matrix, with all
# the implications spelled out, after loading it:
print $matrix->dumpToText(), "\n";
# you can also ask for a list of the emes that match a given object:
print "the vowels are:\n",
join ' ', $matrix->listFeatureClassMembers('VOW');
print "the affricates are:\n",
join ' ', $matrix->listFeatureClassMembers('AFF');
Lingua::FeatureMatrix is a class for managing user-defined
feature-sets. It provides an implementation of datafile parsing that
is generic and useful for anyone defining feature sets of symbols.
If you haven't read the Motivation you might want to skip down to it.
Featuresets are a common way of describing phonetics problems,
e.g. sound change behaviors, but may be useful to people solving other
problems as well. (The included Letter class may, for example, be
useful in writing ligature rules -- if you find this useful for some
other application, please contact the author.)
Users must indicate what type of Eme they are working with. In
fact, users will probably want to define their own. To do this, define
a subclass of Lingua::FeatureMatrix::Eme and indicate that one as
the eme parameter to the new() method call.
Eme typeUsers should not have to provide very much to construct their own
Lingua::FeatureMatrix::Eme that supports all the features you're
interested in.
See Lingua::FeatureMatrix::Eme for details on what's
required to properly subclass Lingua::FeatureMatrix::Eme.
If you'd rather not follow through on all the details specified there,
you can use one of the two stubby subclasses Phone and Letter
provided in the examples/ directory of this distribution as a
jumping-off point. They too are documented, and have a loose licensing
condition for your unrestricted use (see the README).
Takes the following key-value named parameters:
Specifies the desired Lingua::FeatureMatrix::Eme subclass to use
with this Lingua::FeatureMatrix.
TO DO: complete documentation for these methods
To keep this system general, there are several important terms to understand:
TO DO: clarify this vocabulary intro
I use the word eme to describe a single unit (one row of the feature matrix.)
(Think phoneme or grapheme.)
Note these are language-specific. (TO DO: Give example here.)
(Think synchronic rule or feature generalization.)
(Think composite feature.)
(Think single bit of descriptive information.)
You might want to begin by opening the phonematrix.dat file or the
lettermatrix.dat file included in the examples directory of this
distribution. These use the feature sets defined by Phone.pm and
Letter.pm, sample Eme classes each also included in the same
directory.
First, some basic terms that make up the underlying grammar of these datafiles:
Either +, -, or *, indicating the values of 1, 0, and
undef respectively.
A case-sensitive text string like vow indicating the name of the
Eme feature. Always used with SIGN.
A complex grouping of one or more SIGNFEATURE pairs, surrounded
by [], like:
[ +voice +fric +stop ]
a string of characters matching the /\S+/ regular expression. This
is so widely accepting because of the large variety of phonetic
representation schemes available. Leaving this agnostic allows users
to use, e.g.:
(TO DO: include examples here):
Each line in the datafile should be considered an entire statement. You'll find that the datafiles are made up of four kinds of lines. Comments, Eme descriptions, Implicatures, and feature classes. Future versions of this module may include more types of lines.
All lines are insensitive to whitespace, except for a Comment line (which isn't a Comment at all unless there is no whitespace before the '#').
Any line beginning with a '#' is a comment, and the entire line is ignored. Note if the '#' is not the first character on a line, it is not ignored. This is the only place that whitespace is considered in this grammar.
Any line which takes the form:
PHONESYMBOL [ FEATURESET ]
For example,
CH [ +stop +fric -voice ] J [ +stop +fric +voice ] S [ +fric +sib +alv -voice ] Z [ +fric +sib +alv +voice ] SH [ +fric +sib +pal -voice ] ZH [ +fric +sib +pal +voice ] AA [ +low -back -front -tense ] IY [ +high +front +tense ]
It is acceptable, even encouraged, to "underspecify", that is, to
specify only those features which are needed to distinguish each phone
from its neighbors. If you do so, you will probably want to include
extra implicatures though, since any Eme that does not have all
its features specified after the implicatures are processed will
invoke a carp, which can get irritating.
Any line which takes the form
( FEATURESET => FEATURESET )
represents an Implicature. The left FEATURESET is called the
implier and the right is called the implicant.
As a special case, the FEATURESETs involved may omit the [] if
there is only one feature.
Implicatures allow the user to easily encode lots of different Emes
by encoding general "common sense" ideas. For example:
( +stop => +cons )
This means that an Eme that is +stop should be marked +cons
by implication. (If this isn't an obvious implication, you may need
some phonology review, or you may be speaking Czech or Berber, and I
can't help you much with either problem.)
Note that more than one feature may imply the same setting, even to
the same Eme. This is acceptable:
( +fric => +cons ) ( +stop => +cons )
Both of these will apply to the following Eme definition:
CH [ +fric +stop -voice ]
Implicatures are one-way, or else the following wouldn't work:
( -tense => +vow ) ( +tense => +vow )
(The two implications above indicate that if tense is specified
at all, then vow should be + by implication.)
An implicature need not set a single feature in the implicant, nor is it restricted to only one feature in the implier.
( +sib => [ -voice +cons ] ) ( [ +vow +cons ] => [ +glide ] )
Note that some implicatures can point out that a certain field had
better *not* be set (to either plus or minus); here we use the
'ungrammatical' * marker:
( +cons => *tense ) ( +vow => [ *stop *fric ] )
The first example above indicates that if cons is true, then it is
ungrammatical to specify a boolean value for tense, and the
second indicates that if vow is true, then it is ungrammatical
to specify stop or fric. Note that the *fric setting may not
be correct in languages other than English; that's the point of
putting all this in a configuration file.
Sometimes putting "obvious" things into implicatures can help catch silly mistakes in your eme definitions, especially when you can specify ungrammaticality:
# can't be both high and low (though [-high -low] is okay) # seems obvious here... ([+high] => [*low]) ([+low] => [*high]) # 200 lines later, by which point we've forgotten our decision # about the relationship between high and low... # the following eme definition croaks with a warning: EH [ +high +low -tense ] # should have been: # EH [ -high -low -tense ]
Using a * value sets that feature of the Eme to be undef,
rather than 1 or 0, which is Perl's way of indicating "neither
false nor true, but the question is meaningless."
Note that for the time being, the implicatures are applied in the order that they are submitted to the system. Future editions may involve automatic ordering of the implicatures (see Future Improvements).
class AFF => [ +stop +fric ] class LOW_VOW => [ +low +vow ]
I need a tool that constructs objects representing the featureset of a phoneme. The standard linguistic notation for this is (for the 'ch', the 'eh', and the 's' sound in "chess"):
CH [ +stop +fric -voice +palat +cons -vow ] EH [ +vow -cons -low -high +front -tense ] S [ +cons +fric -stop +alv -voice ]
Furthermore, I may want to be able to refer to "feature classes", that is, composite features like "affricate":
class AFF [ +stop +fric ]
(this example would match 'CH' but not 'S' or 'EH').
To complicate things further, the list of primitive features is linguistically controversial, the set of relevant classes varies from language to language, even if you agree on the theoretical primitives, and the choice of symbol set to represent the phoneme (IPA, Sampa, DARPA-bet, etc) is varied and political.
Thus, in the finest Perl sense, TMTOWTDI. The dimensions of flexibility provided are:
You, the user, define what you want to be the featureset by
subclassing Lingua::FeatureMatrix::Eme, distributed with this
module. An added side bonus is that you decide whether the base unit
is a Phone or a Phoneme (or, for that matter, a SoundUnit or
a Letter -- that subclass is your module, and the goal is to
"[put] the focus not so much onto the problem to be solved, but rather
onto the person trying to solve the problem." (see Larry Wall's talk
on Perl and postmodernism http://kiev.wall.org/~larry/pm.html).
You, the user, define what the feature set is, and you define how the phones (er, emes) distribute among those features, using the best of Impatience -- use the existing linguistic typographic conventions, and this module takes care of constructing your objects for you. No translating among conventions for us (that wouldn't be Lazy!).
But let's go one step further. Languages include redundancy, and
sometimes it's boring (and not Lazy) to have to specify yourself
that something that is [+stop] is also [-vow +cons], especially
if you have to specify this for every single [+stop] consonant.
So this module also introduces the concept of an implicature -- you can say, in simple, linguistically-familiar format, that
( [+stop] => [-vow +cons] )
and this will apply for all phones in the current dataset (unless I'm speaking Berber, where this isn't necessarily true...). It's also Lazy, because the module also does the work of letting me know whether I have forgotten to specify any of the features of a given phone:
# probably missing a feature or six; would generate a warning. T [ +cons -vow ]
Along the way, we pick up some Hubris:
Lingua::SoundChange, not to
mention homebrew pronunciation algorithms like Lingua::Soundex.
Original version; created by h2xs 1.21 with options
-CAX Lingua::FeatureMatrix
Now includes lots of error-checking code for handling implicatures better. Still remaining, lots to do, but now can probably be understood by somebody who hasn't read the whole code.
Also includes a lot of documentation, among which is an elaborate Motivation.
If you find any bugs or need additional features, please inform the author -- and check CPAN; this module is under development and may have recently added the feature you need.
For some discussion and ideas about applications of feature matrices:
Jeremy Kahn, <kahn@cpan.org>
Special thanks to Dr. Kate Davis, who acted as the phonetics-theory sounding board for this project.
Includes understanding why the limited test cases provided here fail.
E.g. Lingua::SoundChange.
would require building a graph and toposorting it
But make sure we don't rebuild the wheel.
perl.
| Lingua-FeatureMatrix documentation | Contained in the Lingua-FeatureMatrix distribution. |
package Lingua::FeatureMatrix; use 5.006; use strict; use warnings; use Graph::Directed; use Carp; ################################################################## # package globals our $VERSION = '0.05'; ################################################################## # data methods install using Class::MM use Class::MethodMaker new_with_init => 'new', new_hash_init => 'hash_init', get_set => [ qw [ _emeType _featureClassType ], qw [_eme_new_opts _fclass_new_opts ] ], get_set => [ 'report', 'graph' ], get_set => 'Name', object => [ Graph::Directed => 'implicature_graph',], get_set => 'fh', # object_list => [ Lingua::FeatureMatrix::Implicature => 'implicatures' ], # Lingua::FeatureMatrix::FeatureClass (or subclass) objs hash => 'featureClasses', # Lingua::FeatureMatrix::Eme subclass objs hash => 'emes'; use Lingua::FeatureMatrix::Implicature; use Lingua::FeatureMatrix::FeatureClass; ################################################################## sub init { my $self = shift; my ($class) = ref($self); my (%args) = @_; my $file = $args{file}; if (not defined $file) { croak "$class must be initialized with a 'file' => name", " or 'file' => filehandle key/value pair"; } if (ref($file)) { $self->Name(''); if (UNIVERSAL::isa($file, 'IO::Handle') or UNIVERSAL::isa($file, 'GLOB')) { $self->fh($file); } else { croak "file handed in is a ", ref($file), " and apparently not a descendant of IO::Handle!"; } } else { $self->Name( $file ); require IO::File; # open the filehandle my $fh = IO::File->new($file) or croak "couldn't open $file: $!\n"; $self->fh($fh); } if ($args{report}) { $self->report($args{report}); } else { $self->report(''); } if ($args{graph}) { $self->graph($args{graph}); } else { $self->graph('down'); } $self->implicature_graph( Graph::Directed->new() ); if ($self->graph eq 'down') { $self->implicature_graph->set_attribute(label => 'Feeding and Bleeding ' . 'relationships'); } elsif ($self->graph eq 'up') { $self->implicature_graph->set_attribute(label => 'Possible but ignored feeding ' . 'and bleeding relationships'); } else { warn "don't know what the 'graph' parameter " . $self->graph() . " means\n"; } $self->implicature_graph->set_attribute(ratio => 1); # set up to know which subclasses to use: # we had better find out what Eme class this is $self->_setEmeType($args{eme}, $args{eme_opts}); # we might also have been instructed on a subtype of featureclass to use $self->_setFeatureClass($args{featureclass}, $args{featureclass_opts}); # so far, no requirement to be able to subclass the # implicature. Leave it that way for now. # TO DO: add other more direct ways to initialize dynamically if # needed $self->_loadFile(); #$file); # fill out any features based on implicatures $self->_completeSpecifications(); } ################################################################## sub _setEmeType { my $self = shift; my $class = ref($self); my $emeType = shift; my $eme_new_opts = shift; if (not defined $emeType) { croak "$class must be initialized with an ( " . "'eme' => Lingua::FeatureMatrix::Eme-subclass-name ) key-value pair"; } # make sure that the class specified is loadable # eval "require $emeType"; # if ($@) { # croak "trouble loading $emeType: $@; exiting"; # } if (not $emeType->isa('Lingua::FeatureMatrix::Eme')) { croak "$emeType (provided as 'eme' parameter to $class)" . " is not a Lingua::FeatureMatrix::Eme!\n"; } if ($emeType eq 'Lingua::FeatureMatrix::Eme') { croak "eme parameter to Lingua::FeatureMatrix must be a *derived* " . "subclass, since L::FM::Eme has abstract functions\n"; } if (my $error =$emeType->failsContract()) { croak "$emeType fails to meet Lingua::FeatureMatrix::Eme contract: ", $error, "\n"; } $self->_emeType( $emeType ); # TO DO: undocumented feature allows more powerful subclassing of # Eme objects, by passing more parameters to the Eme-building # routines from this class if (not defined $eme_new_opts) { $self->_eme_new_opts( [] ); } elsif (ref($eme_new_opts) eq 'ARRAY') { # it's an arrayref. that's what we want. $self->_eme_new_opts( $eme_new_opts ); } elsif ( not ref($eme_new_opts) ) { # it's a scalar. Package it up to be an arrayref anyway. $self->_eme_new_opts( [ $eme_new_opts ] ) } else { croak "eme_opts parameter to ", ref($self), " requires arrayref or scalar, not ", ref($eme_new_opts); } } # end _setEmeType ################################################################## sub _setFeatureClass { # TO DO: user-document this undocumented feature, which allows # subclassing of the FeatureClass object. Defaults to the base class # if user ignores it, though. my $self = shift; my $featureclass = shift; my $fClass_opts = shift; if (not defined $featureclass) { $featureclass = 'Lingua::FeatureMatrix::FeatureClass'; } if (not $featureclass->isa('Lingua::FeatureMatrix::FeatureClass')){ croak "featureclass parameter '$featureclass' to ", ref($self), " not a Lingua::FeatureMatrix::FeatureClass " , "subclass!"; } $self->_featureClassType( $featureclass ); if (not defined $fClass_opts) { $self->_fclass_new_opts( [] ); } elsif (ref($fClass_opts) eq 'ARRAY') { $self->_fclass_new_opts( $fClass_opts ); } elsif ( not ref($fClass_opts) ) { $self->_fclass_new_opts( [ $fClass_opts ] ) } else { croak "featureclass_opts parameter to ", ref($self), " requires arrayref or scalar, not ", ref($fClass_opts); } } ################################################################## sub _loadFile { # grabs all lines from config file, strips comments and spaces my $self = shift; my $file = $self->Name(); # open(IN,$file) or die "cannot open $file $!\n"; my $fh = $self->fh(); # sort data into maps, eme-specifications, and class-specifications while (<$fh>) { # clean up lines chomp; $_ = set_utf($_); tr/\x{FEFF}//d; next if /^#/; # drop comments s/\s//g; # and spaces next if not length($_); # and skip blank lines # TO DO: eval the following block, and trap errors. Report them # with line numbers for the user if (/^ \( (.+) \=\> (.+) \) $/x) { # ( +vow => +son ) # ( +cons => *tense ) $self->_readImplicature($1, $2); } elsif (/^class (\S+) \=\> (.+) $/x) { # class AFF => [ +stop +fric ] $self->_readClass($1, $2); } elsif (/^ (\S+) (\[.*) $/x) { # A [ +vow +low -tense ] $self->_readEme($1, $2); } else { # TO DO: document proper .dat file format die "datafile $file has bad format in line $., '$_'\n"; } } close $fh or croak "couldn't close file ", $self->Name(); # close (IN) or die "can't close file $file $!\n"; } ################################################################## sub _readImplicature { my $self = shift; my $class = ref($self); my ($implier, $implicant) = @_; my %implier = $class->_getFeatureSet($implier); my %implicant = $class->_getFeatureSet($implicant); # TO DO: check the featureset used here against whether it's legal. my $implicature = Lingua::FeatureMatrix::Implicature->new(\%implier, \%implicant); $self->add_implicature($implicature); } ################################################################## sub _readClass { my $self = shift; my $class = ref($self); my ($fClassName, $req_features) = @_; my %required = $class->_getFeatureSet($req_features); my $fClassType = $self->_featureClassType(); my $featureClass = $fClassType->new(name => $fClassName, features => \%required); $self->featureClasses( $fClassName => $featureClass ); } ################################################################## sub _readEme { my $self = shift; my ($symbol, $features) = @_; if (defined $self->emes($symbol)) { my $eme = $self->emes($symbol); carp "at line $.: but $symbol previously defined as ", $eme->dumpFeaturesToText($eme->listUserSpecified()); } my $class = ref($self); my (%features) = $class->_getFeatureSet($features); my $eme = $self->_emeType()->new(name => $symbol, options => $self->_eme_new_opts(), %features, ); $self->emes($symbol => $eme); } ################################################################## sub add_implicature { my $self = shift; my Lingua::FeatureMatrix::Implicature $impl = shift; # $self->implicature_graph( my (@otherIndices) = # map {$self->implicature_graph->get_attribute('object', $_)} $self->implicature_graph->vertices(); my $insert_index = scalar (@otherIndices); $self->implicature_graph->add_vertex($insert_index); # add a bunch of details about this object: # machine-readable 'object' $self->implicature_graph->set_attribute(object => $insert_index, $impl); # human-readable 'label' $self->implicature_graph->set_attribute(label => $insert_index, $impl->dumpToText); # see how this new implicature fits into the dependency (ordering). foreach my $otherIdx (@otherIndices) { # yes, we have an n-squared scaling here. Tens of thousands of # rules will have problems, but hundreds should still only take # seconds, maximum. I hope. Don't really know how to build this # graph any other way... :-/ my $other = $self->implicature_graph->get_attribute(object => $otherIdx); my (@inDeps) = $impl->dependsOn($other); my (@outDeps) = $other->dependsOn($impl); if (@inDeps) { if ($self->graph =~ /down/) { $self->implicature_graph->add_edge($otherIdx, $insert_index); $self->implicature_graph->set_attribute( label => $otherIdx, $insert_index, (join " ", @inDeps)); } } if (@outDeps) { # these are dependencies that suggest that it *could* have # bled/fed rules higher up in the ordering. if ($self->graph =~ /up/) { $self->implicature_graph->add_edge($insert_index, $otherIdx); $self->implicature_graph->set_attribute( label => $insert_index, $otherIdx, (join " ", @outDeps)); } if ($self->report() eq 'back-dependencies') { carp $impl->dumpToText(), "(implicature number $insert_index)", " could have been applied before ", $other->dumpToText(), " (implicature number $otherIdx)", " despite their input in the other order."; } } } #end foreach otheridx } ################################################################## sub _completeSpecifications { my $self = shift; # my $ordered_impls = # Lingua::FeatureMatrix::Implicature->order # { # my @ordered = # Lingua::FeatureMatrix::Implicature->sortByRuleOrder($self->implicatures()); # $self->implicatures_clear(); # $self->implicatures_push(@ordered); # } my (@orderedImpls) = $self->orderImplicatures(); foreach my $emeName (sort $self->emes_keys) { my $eme = $self->emes($emeName); # future might consider a toposort here... foreach my $implicature (@orderedImpls) { if ( $implicature->matches( $eme ) ) { $implicature->apply( $eme ); } } my (@missing) = $self->emes($emeName)->listUnspecified(); if (@missing) { warn $self->_emeType . " '$emeName' " . "(a Lingua::FeatureMatrix::Eme subclass) " . "was not fully specified after application of implicatures " . "(missing feature(s) [ @missing ]).\n"; } } } ################################################################# sub _getFeatureSet { # returns a hash of ( feature1 => 1, feature2 => 0, feature3 => undef ) # style data, given "[+feature1 -feature2 *feature3]" style string my $class = shift; my $featureset = shift; if (not defined $featureset) { confess; } $featureset =~ s/^\[//; # remove leading & trailing brackets $featureset =~ s/\]$//; # assumes no features use + or - within their names. my (@featureData) = split( /([+*-])/, $featureset); @featureData = grep {$_ ne ''} @featureData; if ( not (@featureData % 2) ) { # odd } @featureData = map { $_ eq '-' ? 0 : $_ } @featureData; @featureData = map { $_ eq '+' ? 1 : $_ } @featureData; @featureData = map { $_ eq '*' ? (undef $_) : $_ } @featureData; # reversing the @featureData list puts keys before values, instead # of the linguistic standard of value/key # (e.g. +vow becomes vow => 1) # note earlier listings override later return reverse @featureData; } ################################################################## # main public access method. sub matchesFeatureClass { my $self = shift; my ($class) = ref($self); my $symbol = shift; my $className = shift; my $eme = $self->emes($symbol); croak "unrecognized $class symbol $symbol" unless defined $eme; my $featureClass = $self->featureClasses($className); croak "unrecognized class $className" unless defined $featureClass; return $featureClass->matches($eme); } ################################################################## # public (probably debugging) method sub listFeatureClassMembers { my $self = shift; my $className = shift; my $featureClass = $self->featureClasses($className); croak "unrecognized class $className" unless defined $featureClass; my @symbols; foreach my $symbol (sort $self->emes_keys()) { if ( $featureClass->matches( $self->emes($symbol) ) ) { push @symbols, $symbol; } } return @symbols; } ################################################################## sub orderImplicatures { # return "ordered" list of implicatures. my $self = shift; # "By Any Means Necessary" --Malcolm X return map {$self->implicature_graph->get_attribute('object',$_)} sort { $a <=> $b } $self->implicature_graph->vertices(); # Schwartzian Transform is a good means. # map {$_->[0]} # (3) strip sort index # sort { $a->[1] <=> $b->[1] } # (2) sort items by the sort index # map { [$_ => # (1) get a sort index for each # $self->implicature_graph->get_attribute('insert_index', $_)] } # $self->implicature_graph->vertices(); # future improvements will include a toposort call. } ################################################################## sub findEquivalentEmes { my $self = shift; my (@symbols) = $self->emes_keys(); # wantarray returns undef in void context if (not defined wantarray()) { carp "useless call to findEquivalentEmes() in void context"; return; # don't bother doing any work } my %problems; while (@symbols) { my $thisSymbol = shift @symbols; my $thisEme = $self->emes($thisSymbol); foreach my $otherSymbol (@symbols) { if ($thisEme->isEquivalent($self->emes($otherSymbol))) { $problems{$thisSymbol} = $otherSymbol; } # else this eme not equivalent to any remaining eme } # end foreach othersymbol } # end while @symbols remaining if (wantarray()) { return %problems; } else { return (scalar(keys %problems)); } } #end findEquivalentEmes ################################################################## # debugging public method sub dumpToText { # debugging function my $self = shift; my $lineLength = shift; local $Text::Wrap::columns = (defined $lineLength ? $lineLength : $Text::Wrap::columns); my (@text); # loop over emes, dumping each one into filename use Text::Wrap; foreach my $symbol (sort $self->emes_keys) { my $emeText = $self->emes($symbol)->dumpToText(); my $line = $symbol . "\t" . $emeText; push @text, ($Text::Wrap::columns ? wrap ('', "\t ", $line) : $line); } return join ("\n", @text); } ################################################################## sub set_utf { # thanks to perlmonks' grantm # (http://www.perlmonks.org/index.pl?node=grantm) for saving me return pack "U0a*", join '', @_; } ################################################################## 1; __END__