Design and Implementation of a Near-optimal Loglan Syntax

From Lojban
Jump to navigation Jump to search

this file used to live at the late great Planned Languages FTP server, before the WWW was invented.

Design and Implementation of a Near-optimal Loglan Syntax

Submitted by Jeff Prothero, 
May 9, 1990 


A near-optimal surface syntax for an artificial human
language is derived, and a prototype software implementation


The Sapir-Whorf Hypothesis claims that the language
you think in influences the way you think.  Just as you tend to
solve the same problem differently depending on whether you are
coding in ForTran, C, APL or Lisp, so (the Hypothesis claims)
you tend to see the world differently depending on the human
language in which you think:  Your language imprisons your
thoughts.  You cannot escape this prison -- but you can transfer
to another one, by learning a new language.

The loglan (logical language) community is devoted to exploring
the practical implications of the Sapir-Whorf Hypothesis by
constructing and using artificial human languages.  The syntax
of current loglans leaves much to be desired.  The root design
is about thirty-five years old, predating APL, Forth, LALR(1)
parsers, Unix, and much of contemporary formal language theory.
This root design is nearly obscured by accumulated patches.

Current word-resolution algorithms are defined informally in
English, the descriptions run over a dozen pages, and the
algorithm fails on about thirty percent of word boundaries,
at which point inter-word pauses are mandatory.

Current grammars contain over a hundred rules, are context-
sensitive, cannot be recognized in linear time, and after
a third of a century of continuous tinkering, still require
fixes on a weekly basis.  I've designed and implemented
four parsers for various loglans:  It's a pain!

Here I will propose a loglan syntax which:
* Is simple enough to be parsed by a couple of hundred
  lines of straightforward C. (See attached program.)
* Is simple for humans to learn and use.
* Allows for unambiguous resolution of continuous
  human speech.
* Offers near-optimal conciseness and simplicity.


The proposed syntax consists of:
* An alphabet.  "bcdf ghjk lmnp stvz" is suggested, but the
  choice is not critical.
* A pronunciation scheme which makes all sequences of
  letters equally pronouncable, thus decoupling the rest
  of the language design from the details of the human
  vocal tract.
* A set of "affixes" (distinct strings of letters) from
  which to build words.  (These will be used as word
  prefixes, roots, and suffixes.  A word will be the
  concatenation of one or more affixes.)  This affix
  set has the property that affixes of all lengths
  are available, so that frequently used affixes can
  be given a short representation (Zipf's Law).  This
  affix set also has the property that any arbitrary
  string of letters can be decomposed into affixes
  in a most one way, so that arbitrary affixes may
  be concatenated to produce words without introducing
* A distinguished set of affixes which mark word
* A simple (four-rule) grammar which allows the
  expression and parsing of arbitrary syntax
  trees composed of nodes of with zero, one
  or two children.

Problem definition

To make any sort of optimality argument, or indeed any sort
of rational engineering decision, one needs a fairly precise
characterization of the problem to be solved.  We will think
of a language as an encoding used by two agents to exchange
information via a serial channel, which we may think of as a
bitstream, although we will keep in mind that the agents
have human limitations, and that we will want to support
efficient phonetic and written encodings of the bitstream.

We will arbitrarily suppose that these two agents think of
the world as consisting of discrete objects and relations,
and wish to communicate using a finite alphabet of symbols,
rather than (say) modelling the world in terms of smooth
fields and communicating by smoothly varying analog signals.

We have available a number of formally equivalent ways of
representing the knowledge possessed by our two agents.
Suppose one agent knows that John loves Mary.  We may use
the predicate logic and represent this as Loves(John,Mary).
We may use relations, and place a 2-tuple [John, Mary]
in the two-column relation Loves.  Or, we may imagine
John and Mary to be two labelled nodes in a graph, and
Loves to be a labelled, directed edge joining them. Since
we normally think of parsetrees as directed graphs, it will
be convenient to use the graph representation of knowledge
as well:  We will think of the internal knowledge store of
our two communicating agents as consisting of (very) large
graphs with labelled directed edges, some of the nodes
also having labels.  Our agents wish to communicate by
exchanging fragments of these graphs.

Our problem, then is to supply a good encoding scheme
allowing these graph fragments to be linearized, sent
through a bitstream channel, and reconstructed at the
far end.  We require that
1) The reconstruction be unambiguous -- no guesswork.
2) The encoding be as compact as possible -- channel
   bandwidth is considered a precious commodity.
3) The en/decoding system be a simple as possible.
4) The en/decoding system be human-usable -- it must
   cater to the specific strengths and weaknesses of
   human linguistic machinery.

Problem analysis

So far as I can discover, the only reasonable general
plan of attack is as follows:

1) Convert the graph fragment into a tree (or, more
   generally, a forest of trees) by selecting some
   node(s) as tree roots, and then duplicating nodes
   as necessary to satisfy the tree constraint. The
   duplicate nodes are given anaphoric labels so
   the listener will be able to merge them again.
             A             A
            / \           / \
           /   \         /   \
          B     C  -->  B     C
           \   /       /     /
            \ /       /     /
             D       D     D'

2) Linearize the tree by doing a treewalk, introducing
   syntactic markers sufficient to allow unambiguous
   reconstruction of the original tree structure.

            / \
           /   \
          B     C  -->  A( B(D) C(D') )
         /     /
        /     /
       D     D'

3) Encode the linearized tree in the chosen bitstream
   by mapping the abstract symbols to concrete written
   or phonetic encodings.

I'm going to skip Step 1, because I haven't found much
to say about it.  The humanly usable systems are not
very satisfactory from a formal pointer of view --
English uses anaphora like 'it' and 'that' to refer
to previous subexpressions, and depends on pure guess-
work on the listener's part to deduce which subexpression
is intended.  Past loglans have used formally precise
schemes based on mental stacks and counting backward,
schemes which are generally found to be unworkable
in practice.   I don't see anything profound in either
class of system.  It's easy to hack up a suitable instance
of either.

Step two, linearizing the tree, involves basically
putting all the labels on the tree in a row, and
adding just enough syntactic information to allow
later reconstruction of the original tree structure.  
How much information is this?  We will suppose that
nodes in the tree have zero, one, or two children.
(In principle, any tree can be converted to this
form, and in practice human languages seem to
favor this sort of binary syntactic structure.)

For analytic simplicity, consider a strictly binary tree.
Half the nodes are leafs, and half are internal nodes
on the tree.  We cannot possibly reconstruct the tree
without being able to deduce which are which.  A one-bit
tag on each node is exactly necessary and sufficient to
encode the distinction.  Now strip all the leafs off the
tree.  The reduced tree is again half leafs and half
internal nodes, and the argument may be iterated until
only the root node is left.  Thus, half the nodes in
the tree need one bit of structure information, one
quarter of the nodes need two bits of structure
information, one eighth of the nodes need three bits
of structure information, and so forth.  On the
average, a word needs

  Sum (i = 1 to infinity)  0.5  * i    == 2

bits of tree-structure information if the
original tree structure of the sentence is to be
recovered by the listener.

Thus, an optimal linearizing algorithm should add
about two bits of information per node to the
output string.  If it's much more, you're probably
wasting bandwidth, and it it's much less, you're
probably fooling yourself.  Naturally, the information
doesn't necessarily have to be implemented as bit-tags
-- it could be segregated as a prefix string to the
sentence, or whatever.  But it has to provided for
in some fashion.

Step 3, providing a concrete surface representation
for the abstract symbol string:  For analytical
simplicity, we will think of encoding the symbol
string into a bitstream, deferring the task of
encoding the bitstring as inksplats, handwaving,
vocal-cord twanging or whatever.  The human-usability
constraint basically restricts us to one fixed surface
representation for each symbol -- fancy compression
schemes which dynamically compute new representations
for the symbol as a function of the immediately
preceding speechstream are far beyond capabilities
of the human listener.  Thus, the best we can do
on the efficiency front is to assign shorter
codes to frequently-used symbols.  Our unique-
interpretation contraint requires us to choose
a bitstream encoding technique which can be
unambiguously resolved into the intended stream
of symbols by the listener.

So: we need a left-to-right resolvable variable-length
set of bitstring encodings selected on the basis of
symbol frequencies.  Huffman encodings are known to
be an optimal solution to this problem:  Any reasonable
practical system should be either an adaptation of
Huffman coding, or something which can be shown to
be equally simple and effective.

A near-optimal solution

We adopt the alphabet BCDF GHJK LMNP STVZ.  (It is
handy to have the alphabet size be a power of two.
Eight letters would be less concise, thirty-two would
be tough to map onto the standard twenty-six char
character set.  The particular sixteen letters
chosen don't matter.)  To encode an arbitrary
bitstream efficiently, we use these sixteen letters 
as a hex encoding according to the following scheme.
(The capital letters in the right two columns
give the intended pronunciation of each letter
when used as a vowel and when used as a consonant.)

  ------   -----    -----   ---------
    b        0       bEt       Bet
    c        1       shApe     SHape
    d        2       dIp       Dip
    f        3       fOUGHt    Fought

    g        4       gUY       Guy
    h        5       bOOt      THing
    j        6       bOAt      aZure
    k        7       kEEp      Keep

    l        8       REd       THese
    m        9       pREy      Mom
    n       10       pRInt     priNt
    p       11       pROp      Prop

    s       12       tRIte     Site
    t       13       tRUE      True
    v       14       ROver     roVer
    z       15       bREEze    breeZe

Thus we can encode an arbitrary bitstring into
letters by breaking it into groups of four bits,
and replacing each group by the corresponding
letter according to the above table:

   0001 0110 0001 0110 1101 ...
     c    j    b    j    t  ...

By providing both a vowel and a consonant
pronunciation for each letter, and using
them alternately, we can pronounce arbitrary
strings of letters without difficulty.  This is
important:  It modularizes our language design
by decoupling our word-encodings from the
details of the human vocal tract, letting us
concentrate on other issues.  Other than that,
the particular letters and pronunciations
chosen don't matter much, and might be changed
for a non-European audience.  It simplifies
learning to retain alphabetical ordering, of
course, along with the traditional pronunciations
of letters where practical.  The second eight
vowels are simply the first eight with an 'r'
prepended.  With the suggested pronuciations,
the above example "cjbjt" would be pronounced

We will call the smallest semantically meaningful
letter-sequences in our target language "affixes".
These may be complete words, but experience with
previous loglans suggests that it is important
that one allow for internal structure in words --
we should be free construct a word out of one
or more primitive roots, together perhaps with
some prefixes and suffixes.

To group letters into affixes, we know we want a
Huffman-style expanding-opcode sort of scheme. A
simple and effective one is simply to have the
number of leading 1 bits in the affix give the
number of trailing letters in the affix.  This
gives us the following infinite set of affixes,
where '?' may be any single char:

Length 1 (     8 affixes):   b  d  g  j  l  n  s  v
Length 2:(    64 affixes):    c?    h?    m?    t? 
Length 3:(   512 affixes):       f??         p??
Length 4:(  4096 affixes):             k???
Length 5:( 32768 affixes): zb???  zd???  zg???  zj??? 
                           zl???  zn???  zs???  zv??? 
Length 6:(262144 affixes): zc???? zh???? zm???? zt???? 

... with infinitely more to follow, eight times more
for each length.  Since these affixes are based on a
Huffman-style encoding scheme, any arbitrary sequence
of these affixes may be concatenated to form a word,
which word can always be uniquely resolved by the
listener into its constituent affixes.  The listener
would probably simply do this by intuition, being
familiar with the individual affixes, but can also
do this formally and consciously by reasoning, for
example, that any affix beginning with 'm' must
necessarily be exactly two letters long.  Thus,


can only be

    mz d lz g cz j v

This is important, because the listener _must_ be able
to uniquely resolve pauseless speech into its
constituent words and affixes if unambiguous communication
is to be achieved, and we _must_ be able to use arbitrary
sequences of affixes if the language design is to stay
clean and simple.

We just solved the problem of uniquely resolving the
bitstring into affixes:  Now we are faced with the
similar problem of uniquely resolving the resulting
affix-stream into words.  How should this be done?

We have established that words must carry an average
overhead of two bits each of tree-structure
information, so we certainly want our words to average
about eight bits each at a minimum, just to amortize
the overhead decently.  Humans can't reasonably learn
a vocabulary of more than one hundred thousand words,
so the average word length should certainly be less
than seventeen bits.  Eight to sixteen bits per word
is thus the reasonable design goal.

Suppose we mark word boundaries by using a reserved
bitstring.  How many bits long should such a bitstring
be, in order to minimize verbosity?  If we make the
word-boundary marker too long, we will clearly make our
language more verbose.  Less obviously, if we make the word-
boundary marker too short, we will also make our
language more verbose:  By pre-empting valuable short-
word space, we force other frequently-used affixes to use
longer representations than they deserve, thus indirectly
lengthening the average utterance.  The correct length
may be derived from a little elementary information

A bitchannel carries the most information when it
looks completely random.  To the extent that you
can predict incoming bits, they are carrying less
information to you than they could have.  If you
can predict them all exactly, the channel is not
telling you anything at all.  Any language feature
which makes the speechstream look less random is
introducing redundancy, and hence ultimately adding
to the average verbosity of utterances in the
language.  (One may wish to deliberately add
redundant information to a message in order to allow
the listener to detect and correct errors.  This is
a separate issue.)

Applying the above observation to our word-boundary
problem, we should select our word-boundary marker
so that it would chop a truly random bitstream into
"words" with an average length of eight to sixteen bits.

Imagine we divide the bitstream into one-bit codons.
On the average, every other codon will match our
marker, and our words will average     2*1==2 bits.

If we divide the bitstream into two-bit codons,
one codon out of four will match our boundary
marker, and words will average         4*2==8 bits.

If we divide the bitstream into three-bit codons,
one codon out of eight will match our boundary
marker, and words will average         8*3==24 bits.

Thus, we may conclude that whatever syntactic mechanism
we use to clump affixes together to form words should
ideally add about two bits per word to the bitstream.

This is a problem:  Since humans are not terribly
good at shift-and-mask operations, we have set things
up so all affixes are of length 4*N bits.  We don't
have any two-bit affixes!  Luckily, our other problem
provides the solution:  We know we need to attach
about two bits of tree-structure information to
each word to record the tree structure.  Two bits
of end-of-word-indicator affix plus two bits of
tree structure yields a four-bit field -- just
the length of our letters.  Thus our tentative
solution to the word-resolution problem:  A
word is a string of affixes ending with one of the
reserved affixes 'l', 'n' 's' 'v'.  (Any four of the
eight single-letter affixes would do just as well.)
Note that this does _not_ mean that a word ends with
the first 'l', 'n', 's', or 'v' letter!  These letters
will frequently occur inside of multi-letter affixes.
A word is ended only when one of these letters appears
as a single-letter affix.  Example words:

 l      A one-char   word made of the single affix   l
 bl     A two-char   word made of the two    affixes b l
 cln    A three-char word made of the two    affixes cl n
 bdv    A three-char word made of the three  affixes b d v
 fvlczs A five-char  word made of the three  affixes fvl cz s

Assembling words into parsetrees:  The simplest approach
is a precedence grammar.  But instead of assigning each
word a fixed precedence, we will use the word ending to
explicitly specify the precedence for each word:

 Words ending in the affix 'l' are leafs.
 Words ending in the affix 'n' bind most tightly.
 Words ending in the affix 's' bind next most tightly.
 Words ending in the affix 'v' bind next most tightly.

This allows us to express any desired sentence structure
without recourse to parentheses.  (Humans are not very
good at dealing with parentheses, as any Lisp programmer
can attest!)  Thus, using non-leaf words as infix binary
operators for the moment,

    bl cln dl bdv gl  ==  (bl cln dl) bdv gl


    bl clv dl bdl gl  ==  bl clv (dl bdl gl)

just as

    a  *   b   +  c   ==  (a  *   b)  +   c


    a  +   b   *  c   ==  a   +  (b   *   c)

except that we have divorced the meaning of a word
from its precedence and can specify them separately,
whereas conventional mathematical welds the two

In practice, the above four precedence levels provide
enough machinery to resolve most sentence structures.
In principle, however, there is no limit to the number of
precedence levels we might need.  Thus, we amend our
definition of a 'word' to provide an infinite number of
end-of-word affixes:  The alphabetically last four affixes
in each affix length group are end-of-word affixes, and
each such affix binds less tightly than the previous one:

   End-of-word-affix    Precedence
   -----------------    ----------

           l                 0
           n                 1
           s                 2
           v                 3

           ts                4
           tt                5
           tv                6
           tz                7
           pzs               8
           pzt               9
           pzv              10
           pzz              12

          kzzs              13
          kzzt              14
          kzzv              15
          kzzz              16

         zvzzs              17
         zvzzt              18
         zvzzv              19
         zvzzz              20

          ...              ...

This provides us with a pedantically complete system,
but only the first four, or at the very most eight,
affixes are ever likely to be needed in practice.
In the limit of large expressions, this system uses
an average of about 2.6 bits of tree-structure
information per word, compared to the optimal 2.0
bits, incurring a verbosity penalty of about 3-6%
in order to keep things simple for humans.

To distinguish between unary and binary operators, we
borrow a trick from C.  C manages to use '*' as both
a unary and binary operator without incurring ambiguity.
How?  A little thought shows that this works because:
1) C syntax forbids the appearance of two leafs
   without an intervening binary operator.
2) '*' is used only as a prefix unary operator, never
   as a postfix unary operator.  (Mutual exclusion
   is the key -- it would work fine if the reverse
   was true.)
Thus, we can avoid the overhead of explicitly
specifying the arity of a word by adopting a tree
syntax like:



  . . .

:                                         unary_priority_3_expr
| binary_priority_3_expr PRIORITY_3_WORD  unary_priority_3_expr
:                                        binary_priority_2_expr
|                        PRIORITY_3_WORD binary_priority_2_expr

:                                         unary_priority_2_expr
| binary_priority_2_expr PRIORITY_2_WORD  unary_priority_2_expr
:                                        binary_priority_1_expr
|                        PRIORITY_2_WORD binary_priority_1_expr

:                                         unary_priority_1_expr
| binary_priority_1_expr PRIORITY_1_WORD  unary_priority_1_expr
:                                               PRIORITY_0_WORD
|                        PRIORITY_1_WORD        PRIORITY_0_WORD

Thus, using the above grammar

    bl cln bdn gl  ==  bl cln (bdn gl)

just as

    a  *   *   b   ==  a  *   (*   b)

in C.


Let's build a toy language using this syntax.  We'll take
common English words and assign them affixes, then make
up some sentences using this nano-vocabulary.  You can run
these sentences through the supplied parser if you wish.

English word     Toy affix
------------     ---------
   a                b
   and             cd
   be              hk
   but             mt
   can             cn
   car             cc
   do              hd
   drive           mg
   for             tf
   get             ct
   go              mj
   have            hv
   how             hj
   I                g
   if              hf
   in              md
   is              cz
   it              td
   like            tk
   may             mc
   not              d
   of              cv
   on              hn
   or              cj   
   she             ck   
   shoe            ch
   shy             cg
   so              hs
   that            hm
   the             hb
   they            hc
   this            hg
   to              th
   was             hz
   what            ht
   will            ml
   with            hp
   you              j

Sample sentences:

  Gl tkn jl.
  I  like you.

  Ckl tkn gl.
  She likes me.

  Gl mgn hbn ccl.
  I drive the car.

  Gl mgn ckn ccl.
  I drive her car.

  Gl cnn mgn bn ccl.
  I can drive a car.

  Gl tks ckl mgn gn ccl.
  I like her driving my car.

  Gl mln mgn gn ccl thn jl.
  I  will drive my car to you.


We have shown that a simple, near-optimal surface
syntax for a loglan may be obtained by:

1) Adopting an alphabet such as bcdf ghjk lmnp stvz.
2) Using a Huffman-style expanding-opcode technique
   to define a set of affixes of varying lengths.
3) Selecting a subset of these affixes to
   simultaneously mark the end of a word and indicate
   the binding precedence of that word.
4) Using a simple precedence grammar based on unary
   and binary operators.

Compared to existing loglans, such a scheme:
* Is much simpler.  The informal English word-resolution
  algorithm for existing loglan run for a dozen pages
  or more, and the grammars for hundreds of rules.
* Potentially allows for mechanical recognition of
  continuous speech.  Existing loglans require pauses
  between about 30% of the words in a sentence.
* Is suited to laboratory studies of the Sapir-Whorf
  Hypothesis.  Any such studies will require the
  construction of a test language which exhibits some
  property of interest, together with a control
  language which differs from the test language
  only in this property.  A test group will learn
  the test language and a control group the control
  language.  Existing loglans are too complex to
  be quickly built and learned.
* Possesses a certain elegance.
# Recompile planb under Sun Unix (on a 3/160):
cc planb.c -o planb
/*				planb.c					*/
/*  This program parses a simple human language syntax.  It may be      */
/*  compiled with simply "cc planb.c -o planb".                         */

/*                            contents                                  */ 
/*                                                                      */ 
/*      main                                                            */    
/*      findAffixes                                                     */
/*      findParse                                                       */
/*      findWords                                                       */
/*      affix_ends_word TRUE iff affix marks end of word                */
/*      hexVal          Convert bcdf ghjk lmnp stvz to binary value     */
/*      printIndent     Indent a line of pretty-print output            */
/*      printNode       Recursive part of parsetree prettyprint         */ 
/*      printTree       Parsetree prettyprint                           */  
/*      toLower         Convert uppercase letters to lowercase          */
/*                                                                      */ 
/*                            history                                   */ 
/*                                                                      */ 
/* 90Jan17 CrT  Created.                                                */   
/* Width of display, used for prettyprinting: */
#define SCREENwIDTH 80

#define MAIN    1

#define VERSION "1.00 90Jan17"

#define loop	 while (1)

#ifndef TRUE
#define TRUE     1   

#ifndef FALSE
#define FALSE    0   

/* Array to hold pointers to the */ 
/* affixes in the input buffer:  */ 
#define MAXaFFIXES  100 
char *   Affix[ MAXaFFIXES ];  
int      Affix_count;
/* Array to hold pointers to the */ 
/* affixes in the input buffer:  */ 
#define MAXwORDS  100 
char *   Word[            MAXaFFIXES ];  
int      Word_precedence[ MAXaFFIXES ];  
int      Max_precedence_found;
int      Word_count;
int      This_word; 
/* Buffer to read the planb input into: */  
#define MAXiN    10000  
char     Inbuf[      MAXiN ];   
char*    Inbuf_lim;   
/* Buffer to hold the initial parsetree, */ 
/* as a parenthesized string of ints. We */ 
/* also prettyprint the tree into this   */ 
/* buffer, later on in printTree():      */ 
#define MAXoUT   5000 
int      outbuf[    MAXoUT ];   

/* When we've successfully parsed the input, we're left with a giant */  
/* parenthesized expression in outbuf[].  We then construct a proper */ 
/* nodes-and-pointers parsetree using the following node structure:  */ 
#define MAXnODE      100 
struct node {  
    int             word;    /* Index into Word[] array  */ 
    struct node *   kidL;    
    struct node *   kidR;    
} NodeList[ MAXnODE ], *NextNode;    
int Verbose;  

/*      main                                                            */    
main() {   
    struct node *   findParse();
    struct node *   parseTree;
    Verbose = TRUE;
    /* Sign-on message: */   
    printf("\n              -- planb --"              );  
    printf("\n          CurrenT Software's"           );  
    printf("\n      Public-domain LANguage Basher"    );  
    printf("\n         Version %s\n", VERSION         );  
    printf("\nSample sentences you can enter:\n");
    printf("\nGl tkn jl.");
    printf("   (\"I  like you.\")");

    printf("\nCkl tkn gl.");
    printf("   (\"She likes me.\")");

    printf("\nGl mgn hbn ccl.");
    printf("   (\"I drive the car.\")");

    printf("\nGl mgn ckn ccl.");
    printf("   (\"I drive her car.\")");

    printf("\nGl cnn mgn bn ccl.");
    printf("   (\"I can drive a car.\")");

    printf("\nGl tks ckl mgn gn ccl.");
    printf("   (\"I like her driving my car.\")");

    printf("\nGl mln mgn gn ccl thn jl.");
    printf("   (\"I will drive my car to you.\")\n");

    /* Toplevel read-eval-print loop: */    
    loop {     
        printf("\n\n\nEnter sentence:     ");
        if (!*Inbuf) exit(0);

        /* Break the input textstring into affixes: */
        if (!findAffixes())   continue;

        /* Group the affixes into words: */

        /* Group the words into a parsetree: */
        parseTree = findParse();
        if (Verbose)   printTree( parseTree );

/*      findAffixes                                                     */
findAffixes() {
    char *   s;
    char *   d;

    s		= Inbuf;
    /* Kill whitespace, check for unwanted chars */
    /* in input, and fold rest to lower case:    */
    for (s = d = Inbuf;   *s = *d++;   ) {

        /* Convert to lower case: */
        *s = toLower( *s );

        /* Delete blanks, complain about anything else but bcdf ghjk lmnp stvz: */
        switch (*s) {

        case 'b':   case 'c':   case 'd':   case 'f':
        case 'g':   case 'h':   case 'j':   case 'k':
        case 'l':   case 'm':   case 'n':   case 'p':
        case 's':   case 't':   case 'v':   case 'z':

        case ' ':
        case '\t':

            printf("Invalid char '%c'\n", *s );
            return FALSE;
    if (Verbose)   printf("\nFolded   input: '%s'\n",Inbuf);

    /* Insert a blank after every char in input string: */
    d     = &Inbuf[ 2*(s-Inbuf) ];    /* Set d twice as far into Inbuf as s.   */
    *d--  = *s--;                     /* Deposit new terminal nul.             */
    while (d > Inbuf)     {   *d-- = ' ';   *d-- = *s--;   }

    /* Break input into affixes, following rule that number of       */
    /* trailing chars in affix is given by number of leading 1 bits: */
    s = d = Inbuf;    /* Actually, this is already true. */   
    Affix_count = 0;
    while (*s) {

        int hex = hexVal( *s );
        int leading_1s = 0;

        /* Remember start of affix, and keep count of affixes found: */
        Affix[ Affix_count++ ] = d;

        /* Count number of leading 1s: */
        while (hex == 0xF) {
            leading_1s += 4;
            *++d = *(s+=2);
            if (!*s) { printf("Input ends with incomplete affix"); return FALSE; }
            hex = hexVal( *s );
        while (hex & 1)     {   hex >>= 1;   ++leading_1s;   }

        /* Eat that number of trailing chars: */
        while (leading_1s--) {
            *++d = *(s+=2);
            if (!*s) { printf("Input ends with incomplete affix"); return FALSE; }

        /* Lay down terminal nul for affix: */
        *++d = '\0';

        /* Bump s to start of next affix: */
        *++d = *(s+=2);

    if (Verbose) {
        printf("\nAffixes  found:");
            int i;
            for (i = 0;   i < Affix_count;   ++i)     printf(" %s",Affix[i]);

    /* Remember first free location in Inbuf[]: */
    Inbuf_lim = d;

/*      findParse                                                       */
struct node *   findParse_unary(  precedence )
int                               precedence;
    struct node *   findParse_binary();
    struct node *   nod;

    /* If we're out of words or don't have a word of     */
    /* correct precedence, just return right expression: */
    if (This_word == Word_count                         ||
        Word_precedence[ This_word ]   !=   precedence
    ) {
        return   findParse_binary( precedence-1 );
    /* Build and return a parsetree node for our word: */
    nod       = NextNode++;
    nod->word = This_word++;
    nod->kidL = (struct node *) FALSE;
    nod->kidR = findParse_unary( precedence );
    return nod;
struct node *   findParse_binary( precedence )
int                               precedence;
    struct node *   left;
    struct node *   nod ;

    /* Test for limiting case of recursion: */
    if (precedence < 0)            return (struct node *) FALSE;

    /* Parse left subexpression: */
    left = findParse_unary( precedence );

    /* If we're out of words or don't have a word of    */
    /* correct precedence, just return left expression: */
    while (This_word < Word_count                       &&
        Word_precedence[ This_word ]   ==   precedence
    ) {
        /* Build and return a parsetree node for our word: */
        nod       = NextNode++;
        nod->word = This_word++;
        nod->kidL = left;
        nod->kidR = findParse_unary( precedence );
        left      = nod;
    return left;
struct node *   findParse() {
    /* Here we group the words into a parsetree.  The parse is based on  */
    /* the idea that priority-0 words are leafs, and the other words are */
    /* unary or binary operators. Whether such words are unary or binary */
    /* is determined from context, using a grammar much like that for    */
    /* the '*', '-' and '&' unary/binary operators of C.  The grammar is */
    /* technically infinite since we have an infinite number of possible */
    /* precedences.  Here it is in YACC syntax:                          */
    /*                                                                   */
    /*   . . .                                                           */
    /* %token PRIORITY_3_WORD                                            */
    /* %token PRIORITY_2_WORD                                            */
    /* %token PRIORITY_1_WORD                                            */
    /* %token PRIORITY_0_WORD                                            */
    /*                                                                   */
    /* %%                                                                */
    /*                                                                   */
    /*   . . .                                                           */
    /*                                                                   */
    /* binary_priority_3_expr                                            */
    /* : unary_priority_3_expr                                           */
    /* | unary_priority_3_expr PRIORITY_3_WORD  binary_priority_3_expr   */
    /* ;                                                                 */
    /* unary_priority_3_expr                                             */
    /* :                                        binary_priority_2_expr   */
    /* |                       PRIORITY_3_WORD   unary_priority_3_expr   */
    /* ;                                                                 */
    /*                                                                   */
    /* binary_priority_2_expr                                            */
    /* : unary_priority_2_expr                                           */
    /* | unary_priority_2_expr PRIORITY_2_WORD  binary_priority_2_expr   */
    /* ;                                                                 */
    /* unary_priority_2_expr                                             */
    /* :                                        binary_priority_1_expr   */
    /* |                       PRIORITY_2_WORD   unary_priority_2_expr   */
    /* ;                                                                 */
    /*                                                                   */
    /* binary_priority_1_expr                                            */
    /* : unary_priority_1_expr                                           */
    /* | unary_priority_1_expr PRIORITY_1_WORD  binary_priority_1_expr   */
    /* ;                                                                 */
    /* unary_priority_1_expr                                             */
    /* :                                               PRIORITY_0_WORD   */
    /* |                       PRIORITY_1_WORD   unary_priority_1_expr   */
    /* ;                                                                 */

    /* Move all the parsetree nodes to the freelist: */
    NextNode = NodeList;

    /* Reset global current-word pointer to start of Word[]: */
    This_word = 0;

    /* Now a trivial recursive-descent parse does the trick: */
    return   findParse_binary( Max_precedence_found );

/*      findWords                                                       */
findWords() {

    /* Here we fill Words[] and Word_precedence[] with the text string */
    /* and precedence respectively of each word found in the input:    */
    char *   s;
    char *   d;
    int      a = 0;

    Word_count              = 0;
    Max_precedence_found    = 0;

    /* We will write the words we find into the free space at the end of Inbuf: */
    d          = Inbuf_lim;
    /* While unprocessed affixes remain: */
    while (a < Affix_count) {

        int precedence  = 0;
        int end_of_word = affix_ends_word( Affix[a], &precedence );
        /* Remember start of word, and keep count of words found: */
        Word[ Word_count ] = d;

        /* Copy affix to where we're constructing our word: */
        for (s = Affix[a];   *d++ = *s++;   );
        /* Keep appending affixes until we read end of word or input: */
        while (!end_of_word   &&   a < Affix_count) {
            d[-1]       = '-';
            for (s = Affix[a];   *d++ = *s++;   );
            end_of_word = affix_ends_word( Affix[a], &precedence );

        /* Remember precedence of the word: */
        Word_precedence[ Word_count++ ] = precedence;

        /* Remember maximum precedence found: */
        if (Max_precedence_found < precedence)   Max_precedence_found = precedence;

    if (Verbose) {
        int p;
        printf("\nWords found, by precedence:\n");
        for (p = Max_precedence_found;   p >= 0;   --p) {
            int w;
            printf(   "Precedence %3d:",   p   );
            for (w = 0;   w < Word_count;   ++w) {
                if (Word_precedence[w] == p)     printf(   " %s",   Word[ w ]   );
                else {
                    int i = strlen( Word[w] ) +1;
                    while (i--)   printf(" ");

    return TRUE;

/*      affix_ends_word      TRUE iff affix marks end of word           */
affix_ends_word( affix, precedence )
char *           affix;                /* input */
int  *                  precedence;    /* valid only if fn returns TRUE */
    if (!strcmp( affix,     "l" )) { *precedence =  0;   return TRUE; }
    if (!strcmp( affix,     "n" )) { *precedence =  1;   return TRUE; }
    if (!strcmp( affix,     "s" )) { *precedence =  2;   return TRUE; }
    if (!strcmp( affix,     "v" )) { *precedence =  3;   return TRUE; }

    if (!strcmp( affix,    "ts" )) { *precedence =  4;   return TRUE; }
    if (!strcmp( affix,    "tt" )) { *precedence =  5;   return TRUE; }
    if (!strcmp( affix,    "tv" )) { *precedence =  6;   return TRUE; }
    if (!strcmp( affix,    "tz" )) { *precedence =  7;   return TRUE; }

    if (!strcmp( affix,   "pzs" )) { *precedence =  8;   return TRUE; }
    if (!strcmp( affix,   "pzt" )) { *precedence =  9;   return TRUE; }
    if (!strcmp( affix,   "pzv" )) { *precedence = 10;   return TRUE; }
    if (!strcmp( affix,   "pzz" )) { *precedence = 11;   return TRUE; }

    if (!strcmp( affix,  "kzzs" )) { *precedence = 12;   return TRUE; }
    if (!strcmp( affix,  "kzzt" )) { *precedence = 13;   return TRUE; }
    if (!strcmp( affix,  "kzzv" )) { *precedence = 14;   return TRUE; }
    if (!strcmp( affix,  "kzzz" )) { *precedence = 15;   return TRUE; }

    if (!strcmp( affix, "zvzzs" )) { *precedence = 16;   return TRUE; }
    if (!strcmp( affix, "zvzzt" )) { *precedence = 17;   return TRUE; }
    if (!strcmp( affix, "zvzzv" )) { *precedence = 18;   return TRUE; }
    if (!strcmp( affix, "zvzzz" )) { *precedence = 19;   return TRUE; }

    if (!strcmp( affix,"ztzzzs" )) { *precedence = 20;   return TRUE; }
    if (!strcmp( affix,"ztzzzt" )) { *precedence = 21;   return TRUE; }
    if (!strcmp( affix,"ztzzzv" )) { *precedence = 22;   return TRUE; }
    if (!strcmp( affix,"ztzzzz" )) { *precedence = 23;   return TRUE; }

    /* There are infinitely more (the last 4 numerically of each affix size) */
    /* but why be pedantic?  Only the first eight have any real chance of    */
    /* being used.                                                           */

    return FALSE;

/*      hexVal          Convert bcdf ghjk lmnp stvz to binary value     */
hexVal( c )
int     c;
    switch (c) {

    case 'b':    return 0x0;
    case 'c':    return 0x1;
    case 'd':    return 0x2;
    case 'f':    return 0x3;

    case 'g':    return 0x4;
    case 'h':    return 0x5;
    case 'j':    return 0x6;
    case 'k':    return 0x7;

    case 'l':    return 0x8;
    case 'm':    return 0x9;
    case 'n':    return 0xA;
    case 'p':    return 0xB;

    case 's':    return 0xC;
    case 't':    return 0xD;
    case 'v':    return 0xE;
    case 'z':    return 0xF;

        printf(   "hexVal: internal error '%c'",   c   );
/*      printIndent     Indent a line of pretty-print output            */
char *   printIndent( depth, out )
int		      depth;
char *			     out;
    int   col;

    *out++ = '\n';
    for (col = 0;   col < 2*depth;   ++col)   *out++ = ' ';

    return out;

/*      printNode       Recursive part of parsetree prettyprint         */ 
char* printNode( nod, depth, out0 )
struct node *	 nod;
int		      depth;
char *			     out0;
    static char *   lParen = "<([{";
    static char *   rParen = ">)]}";
    int             i;
    char *          out = out0;

    if (!nod)	 return out;

    /* Leaf node? */
    if (!nod->kidL   &&   !nod->kidR) {

	/* Yes: */
        sprintf( out, "%s ", Word[ nod->word ]  );
	out += strlen( out );
        return out;

    /* Internal node.  Try fitting it all on one line: */

    /* Open clause: */
    sprintf( out,  "%c%s: ",   lParen[ depth & 3 ],   Word[ nod->word ]   );
    out += strlen( out );

    /* Do kids: */
    out = printNode( nod->kidL, depth+1, out ); 
    out = printNode( nod->kidR, depth+1, out ); 

    /* Close clause: */
    sprintf( out,  "%c ",      rParen[ depth & 3 ]   );
    out += strlen( out );

    /* Did it all fit on one line? */
    if ((out - out0) + 2*depth	 >=   SCREENwIDTH) {

	/* No, break it up into multiple lines: */

	/* Erase previous one-line try: */
        out = out0;

        /* Open clause: */
	sprintf( out,  "%c%s: ",   lParen[ depth & 3 ],   Word[ nod->word ]   );
	out += strlen( out );

        /* Do kids: */
        if (nod->kidL) {
            out = printIndent( depth, out );
	    out = printNode( nod->kidL, depth, out );
        if (nod->kidR) {
            out = printIndent( depth, out );
	    out = printNode( nod->kidR, depth, out );

	/* Close clause: */
	out = printIndent( depth, out );
	sprintf( out,  "%c ",	   rParen[ depth & 3 ]	 );
	out += strlen( out );

    return out;

/*      printTree       Parsetree prettyprint                           */  
printTree(	self )
struct node *   self; 
    printf( "\nParse:    " );
    printNode( self,  5, (char*)outbuf );
    puts(		 (char*)outbuf );
/*      toLower         Convert uppercase letters to lowercase          */
toLower( c )
int	 c;
    if (c < 'A'   ||   c > 'Z')   return c;

    return   c	+  ('a' - 'A');

--Submitted by Jeff Prothero,
  May 9, 1990

 -------------------- cut here --------------------