fst-compiler-utf8(1) Two compilers for SFST programs


fst-compiler grammar-file [ output-file ]
fst-compiler-utf8 grammar-file [ output-file ]


Store the transducer in compact format which is used by fst-infl2 and.
Store the transducer in lowmem format.
Switch upper and lower level of the transducer. You have to use this switch in order to use fst-infl (fst-infl2, fst-infl3) for generation rather than analysis.


fst-compiler is a compiler for finite-state transducer programs. It generates a minimized finite state transducer which can be used with fst-mor, fst-infl, fst-print, fst-compare, fst-parse, and fst-lattice. The compact transducer representation which is generated with the -c flag, is supported by fst-infl2, fst-train, and fst-match. The memory-efficient transducer representation which is generated with the -l flag, is only supported by fst-infl3.

The first program argument is the name of a file which contains the transducer program. The programming language is described below. The second argument is the name of the file to which the resulting transducer will be written in binary form. If a second argument is missing, the output will be written to stdout.

fst-compiler-utf8 differs from fst-compiler only in the character encoding. fst-compiler-utf8 supports UTF8 encoding of the source files whereas fst-compiler is to be used for 8-Bit character codes like latin1 which are an extension of the ASCII code. Information about the encoding is stored in the transducer files and used by the other SFST programs.


A transducer program consists of an (optional) sequence of alphabet and variable definitions followed by a single transducer expression which defines the result transducer.


An alphabet definition consists of the keyword ALPHABET followed by = and some transducer expression e.g.

    ALPHABET = [a-z]:[A-Z]

This command redefines the alphabet as the set of symbol pairs occurring on the transitions of the transducer. Occurrences of two-level operators, negation operators and unquoted periods always have to be preceded by an alphabet definition.


There are two different types of variables. Symbol set variables are enclosed by hash signs (#) and take symbol sequences (see below) as values:

    #UC# = A-Z

    #LC# = a-z

Transducer variables are enclosed by dollar signs and take transducer expressions as values:

    $MAP$ = [a-z]:[A-Z]+

    $MAP$ = [#LC#]:[#UC#]+

Variables whose name starts with the symbol `=' are special agreement variables. If an agreement variable occurs more than once in a transducer expression, it will always have the same value. Consider the following transducer program:

    $=1$ = [abc]

    $=1$ X $=1$

The result transducer recognizes the strings aXa, bXb, and cXc. Only acyclic transducers (i.e. transducers with a finite set of string mappings) can be assigned to agreement variables.


A symbol is either

- a single character like A s 5,

- a quoted character like \* or \_,

- a multi-character symbol like <X> or <ab.c5> (which is always
enclosed in angle brackets) or
- a backslash followed by a number which is the numeric code of the
designated character

- the null symbol <>.

Symbol sequence

A symbol sequence is a sequence of characters, multi-character symbols and character ranges, e.g. a-z \. <x>.

symbol range

A symbol range is either

- a single symbol

- a symbol sequence enclosed in square brackets like [A-Za-z] or

- a symbol sequence starting with ^ and enclosed in square brackets like [^A-Za-z] (designating the complement of [a-zA-Z]) or

- the period (which represents any symbol from the alphabet)

Transducer expressions

A transducer expression (TE) is recursively defined as follows:

- A pair of two symbol ranges separated by a colon is a TE.


- A single symbol range like [a-z] is a TE.
It is a short form for [a-z]:[a-z].
- Two symbol sequences enclosed in braces and separated by a colon are
a TE. {a[bc]}:{def} is equivalent to a:d b:e <>:f | a:d c:e <>:f.
- X Y is a TE if X and Y are TEs.
(Blanks are ignored unless they are quoted.)
- (X) is a TE if X is a TE.
- X op is a TE is X is a TE and op is either * (Kleene's star operator), +
(Kleene's plus operator), or ? (optionality operator)
- op X is a TE is X is a TE and op is either ! (negation operator), ^
(target language extraction operator), _ (source language extraction operator), or ^_ (source and target switch operator).
- X op Y is a TE is X and Y are TEs and op is either & (conjunction
operator), | (disjunction operator), || (composition operator), or - (subtraction operator)
- L x op y R is a TE if L and R are TEs, x and y are symbol ranges and
op is either => (two-level restriction), <= (two-level coercion), or <=> (two-level restriction and coercion).
- X op L__R is a TE if X, L and R are TEs and op is either ^-> (upward
replacement), _-> (downward replacement), /-> (leftward replacement) or \-> (rightward replacement). Furthermore, L and R must define automata (i.e. which map their strings onto themselves). These operators correspond to Karttunen's replace operators. If the arrow is followed by a question mark (?), the replacement becomes optional.
- X << l is a TE if X is a TE, and l is either of the form
a or the form a:b where a and b are single characters or symbols. The result is a transducer where l was freely inserted into X. The transducer ab << c for instance is equivalent to c*ac*bc*.
- X op Y L1__R2, ... , LN__RN is a TE if X,Y, L1 through LN and R1
through RN are TEs, and op is either => (general restriction), <= (general coercion), ^=> (general surface restriction), ^<= (general surface coercion), ^<=> (general surface restriction and coercion), _=> (general deep restriction), _<= (general deep coercion), _<=> (general deep restriction and coercion). (These operators were implemented following a suggestion by Anssi Yli-Jyra.)
- "fname" is a TE. The compiler reads the file named fname and turns
it into a transducer of the form line1|line2|line3|... where linex is the x-th line of the file. All characters other than : and \ are interpreted literally (i.e. not as operators). This TE is typically used e.g. to read morpheme list from a file.
- "<fname>" is a TE. The compiler reads a pre-compiled transducer from
the file named fname. This

Further Features

Comments start with the symbol % and extend up to the end of the line. Blanks are ignored unless they are quoted. Expressions terminate at the end of a line unless the end of line is preceded by a backslash. The command

#include "fname"

can be used to insert source code from a file named fname. The command

RE >> "fname"

stores the regular expression RE in the file fname.


Here is an example of a simple transducer program. Assuming that the file "adj-stems" contains the two lines


this transducer will correctly analyse the adjective forms easy, easier, easiest and late, later, and latest.

ALPHABET = [a-zA-Z] y:i e:<> <ADJ>:<>

$R$ = y<=>i (<ADJ>:<> e)

$R2$ = e<=><> (<ADJ>:<> e)

$R$ = $R$ & $R2$

$Stems$ = "adj-stems"

$S$ = $Stems$ <ADJ> (<pos>:<>|<cmp>:{er}|<sup>:{est})

$S$ || $R$


fst-compiler returns 0 unless some error occurs.


The compiler gets the operator precedence wrong in case of two-level rules and interprets the expression "ab c<=>d ef" as "a(b c<=>d (ef))". Therefore, you should always surround the left context of two-level rules with parenthesis: (ab) c<=>d (ef)


Helmut Schmid, Institute for Computational Linguistics, University of Stuttgart, Email: [email protected], This software is available under the GNU Public License.