rdf_litindex.pl

/*  Part of SWI-Prolog

    Author:        Jan Wielemaker
    E-mail:        J.Wielemaker@vu.nl
    WWW:           http://www.swi-prolog.org
    Copyright (c)  2006-2020, University of Amsterdam
                              VU University Amsterdam
                              CWI, Amsterdam
    All rights reserved.

    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions
    are met:

    1. Redistributions of source code must retain the above copyright
       notice, this list of conditions and the following disclaimer.

    2. Redistributions in binary form must reproduce the above copyright
       notice, this list of conditions and the following disclaimer in
       the documentation and/or other materials provided with the
       distribution.

    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
    FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
    COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
    INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
    BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
    LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
    CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
    LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
    ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
    POSSIBILITY OF SUCH DAMAGE.
*/

:- module(rdf_litindex,
          [ rdf_set_literal_index_option/1,     % +Options
            rdf_tokenize_literal/2,             % +Literal, -Tokens
            rdf_find_literal/2,                 % +Spec, -Literal
            rdf_find_literals/2,                % +Spec, -ListOfLiterals
            rdf_token_expansions/2,             % +Spec, -Expansions
            rdf_stopgap_token/1,                % -Token

            rdf_literal_index/2,                % +Type, -Index
            rdf_delete_literal_index/1          % +Type
          ]).
:- autoload(rdf_db,
	    [ rdf_keys_in_literal_map/3,
	      rdf_find_literal_map/3,
	      rdf_new_literal_map/1,
	      rdf_monitor/2,
	      rdf_current_literal/1,
	      rdf_reset_literal_map/1,
	      rdf_insert_literal_map/4,
	      rdf_delete_literal_map/2,
	      rdf/3,
	      rdf_delete_literal_map/3,
	      rdf_insert_literal_map/3,
	      rdf_statistics_literal_map/2
	    ]).
:- autoload(library(apply),[maplist/3]).
:- use_module(library(debug),[debug/3]).
:- autoload(library(double_metaphone),[double_metaphone/2]).
:- autoload(library(error),
	    [instantiation_error/1,must_be/2,domain_error/2]).
:- autoload(library(lists),[member/2,flatten/2,append/3]).
:- autoload(library(porter_stem),[tokenize_atom/2]).
:- autoload(library(snowball),[snowball/3]).

/** <module> Search literals

This module finds literals of the RDF  database based on words, stemming
and _sounds like_ (metaphone).  The normal user-level predicate is

  - rdf_find_literals/2
*/

:- dynamic
    literal_map/2,                  % Type, -Map
    map_building/2,                 % Type, -Queue
    new_token/2,                    % Hook
    setting/1,
    stopgap/1.
:- volatile
    literal_map/2.
:- multifile
    tokenization/2,                 % +Literal, -Tokens
    exclude_from_index/2.           % +Which, +Token


setting(verbose(false)).                % print progress messages
setting(index_threads(1)).              % # threads for creating the index
setting(index(thread(1))).              % Use a thread for incremental updates
setting(stopgap_threshold(50000)).      % consider token a stopgap over N

%!  rdf_set_literal_index_option(+Options:list)
%
%   Set options for the literal package.  Currently defined options
%
%           * verbose(Bool)
%           If =true=, print progress messages while building the
%           index tables.
%
%           * index_threads(+Count)
%           Number of threads to use for initial indexing of
%           literals
%
%           * index(+How)
%           How to deal with indexing new literals.  How is one of
%           =self= (execute in the same thread), thread(N) (execute
%           in N concurrent threads) or =default= (depends on number
%           of cores).
%
%           * stopgap_threshold(+Count)
%           Add a token to the dynamic stopgap set if it appears in
%           more than Count literals.  The default is 50,000.

rdf_set_literal_index_option([]) :- !.
rdf_set_literal_index_option([H|T]) :-
    !,
    set_option(H),
    rdf_set_literal_index_option(T).
rdf_set_literal_index_option(Option) :-
    set_option(Option).

set_option(Term) :-
    check_option(Term),
    functor(Term, Name, Arity),
    functor(General, Name, Arity),
    retractall(setting(General)),
    assert(setting(Term)).

check_option(X) :-
    var(X),
    !,
    instantiation_error(X).
check_option(verbose(X)) :-
    !,
    must_be(boolean, X).
check_option(index_threads(Count)) :-
    !,
    must_be(nonneg, Count).
check_option(stopgap_threshold(Count)) :-
    !,
    must_be(nonneg, Count).
check_option(index(How)) :-
    !,
    must_be(oneof([default,thread(_),self]), How).
check_option(Option) :-
    domain_error(literal_option, Option).


                 /*******************************
                 *            QUERY             *
                 *******************************/

%!  rdf_find_literal(+Spec, -Literal) is nondet.
%!  rdf_find_literals(+Spec, -Literals) is det.
%
%   Find literals in the RDF database matching Spec.  Spec is defined
%   as:
%
%   ==
%   Spec ::= and(Spec,Spec)
%   Spec ::= or(Spec,Spec)
%   Spec ::= not(Spec)
%   Spec ::= sounds(Like)
%   Spec ::= stem(Like)             % same as stem(Like, en)
%   Spec ::= stem(Like, Lang)
%   Spec ::= prefix(Prefix)
%   Spec ::= between(Low, High)     % Numerical between
%   Spec ::= ge(High)               % Numerical greater-equal
%   Spec ::= le(Low)                % Numerical less-equal
%   Spec ::= Token
%   ==
%
%   sounds(Like) and stem(Like) both map to  a disjunction. First we
%   compile the spec to normal form:   a disjunction of conjunctions
%   on elementary tokens. Then we execute   all the conjunctions and
%   generate the union using ordered-set algorithms.
%
%   Stopgaps are ignored. If the final result is only a stopgap, the
%   predicate fails.
%
%   @tbd Exploit ordering of numbers and allow for > N, < N, etc.

rdf_find_literal(Spec, Literal) :-
    rdf_find_literals(Spec, Literals),
    member(Literal, Literals).

rdf_find_literals(Spec, Literals) :-
    compile_spec(Spec, DNF),
    DNF \== @(stopgap),
    token_index(Map),
    lookup(DNF, Map, _, SuperSet),
    flatten(SuperSet, Set0),
    sort(Set0, Literals).

%!  rdf_token_expansions(+Spec, -Extensions)
%
%   Determine which extensions of  a   token  contribute  to finding
%   literals.

rdf_token_expansions(prefix(Prefix), [prefix(Prefix, Tokens)]) :-
    token_index(Map),
    rdf_keys_in_literal_map(Map, prefix(Prefix), Tokens).
rdf_token_expansions(sounds(Like), [sounds(Like, Tokens)]) :-
    metaphone_index(Map),
    rdf_find_literal_map(Map, [Like], Tokens).
rdf_token_expansions(stem(Like), [stem(Like, Tokens)]) :-
    stem_index(Map),
    rdf_find_literal_map(Map, [Like], Tokens).
rdf_token_expansions(Spec, Expansions) :-
    compile_spec(Spec, DNF),
    token_index(Map),
    lookup(DNF, Map, SCS, _),
    flatten(SCS, CS),
    sort(CS, Expansions0),
    join_expansions(Expansions0, Expansions).

join_expansions([], []).
join_expansions([H0|T0], [H|T]) :-
    untag(H0, Tag, V0),
    Tag =.. L0,
    append(L0, [[V0|Values]], L1),
    H =.. L1,
    join_expansions_by_tag(T0, Tag, T1, Values),
    join_expansions(T1, T).

join_expansions_by_tag([H|T0], Tag, T, [V0|VT]) :-
    untag(H, Tag, V0),
    !,
    join_expansions_by_tag(T0, Tag, T, VT).
join_expansions_by_tag(L, _, L, []).

lookup(@(false), _, [], []) :- !.
lookup(or(H0,T0), Map, [CH|CT], [H|T]) :-
    !,
    lookup(H0, Map, CH, H),
    lookup(T0, Map, CT, T).
lookup(H0, Map, [C], [H]) :-
    lookup1(H0, Map, C, H).

lookup1(Conj, Map, Cond, Literals) :-
    phrase(conj_to_list(Conj), List),
    !,
    rdf_find_literal_map(Map, List, Literals),
    (   Literals \== []
    ->  phrase(conj_to_cond(Conj), Cond)
    ;   Cond = []
    ).
lookup1(_, _, _, []).

conj_to_list(and(A,B)) -->
    !,
    conj_to_list(A),
    conj_to_list(B).
conj_to_list(@(false)) -->
    !,
    {fail}.
conj_to_list(Tagged) -->
    { untag(Tagged, L) },
    !,
    [L].
conj_to_list(L) -->
    [L].


conj_to_cond(and(A,B)) -->
    !,
    conj_to_cond(A),
    conj_to_cond(B).
conj_to_cond(Tagged) -->
    { untag(Tagged, _) },
    !,
    [ Tagged ].
conj_to_cond(_) -->
    [].


%!  compile_spec(+Spec, -Compiled)
%
%   Compile a specification as above into disjunctive normal form

compile_spec(Spec, DNF) :-
    expand_fuzzy(Spec, Spec2),
    nnf(Spec2, NNF),
    dnf(NNF, DNF).


expand_fuzzy(Var, _) :-
    var(Var),
    !,
    throw(error(instantiation_error, _)).
expand_fuzzy(sounds(Like), Or) :-
    !,
    (   atom(Like)
    ->  metaphone_index(Map),
        double_metaphone(Like, Key),
        rdf_find_literal_map(Map, [Key], Tokens),
        list_to_or(Tokens, sounds(Like), Or)
    ;   expand_fuzzy(Like, Or)
    ).
expand_fuzzy(stem(Like), Or) :-
    !,
    expand_fuzzy(stem(Like, en), Or).
expand_fuzzy(stem(Like, Lang), Or) :-
    !,
    (   atom(Like)
    ->  stem_index(Map),
        stem(Like, Lang, Key),
        rdf_find_literal_map(Map, [Key], Tokens),
        list_to_or(Tokens, stem(Like), Or)
    ;   expand_fuzzy(Like, Or)
    ).
expand_fuzzy(prefix(Prefix), Or) :-
    !,
    (   atom(Prefix)
    ->  token_index(Map),
        rdf_keys_in_literal_map(Map, prefix(Prefix), Tokens),
        list_to_or(Tokens, prefix(Prefix), Or)
    ;   expand_fuzzy(Prefix, Or)
    ).
expand_fuzzy(case(String), Or) :-
    !,
    (   atom(String)
    ->  token_index(Map),
        rdf_keys_in_literal_map(Map, case(String), Tokens),
        list_to_or(Tokens, case(String), Or)
    ;   expand_fuzzy(String, Or)
    ).
expand_fuzzy(or(A0, B0), E) :-
    !,
    expand_fuzzy(A0, A),
    expand_fuzzy(B0, B),
    simplify(or(A,B), E).
expand_fuzzy(and(A0, B0), E) :-
    !,
    expand_fuzzy(A0, A),
    expand_fuzzy(B0, B),
    simplify(and(A,B), E).
expand_fuzzy(not(A0), not(A)) :-
    !,
    expand_fuzzy(A0, A).
expand_fuzzy(between(Low, High), Or) :-
    !,
    token_index(Map),
    rdf_keys_in_literal_map(Map, between(Low, High), Tokens),
    list_to_or(Tokens, between(Low, High), Or).
expand_fuzzy(le(High), Or) :-
    !,
    token_index(Map),
    rdf_keys_in_literal_map(Map, le(High), Tokens),
    list_to_or(Tokens, le(High), Or).
expand_fuzzy(ge(Low), Or) :-
    !,
    token_index(Map),
    rdf_keys_in_literal_map(Map, ge(Low), Tokens),
    list_to_or(Tokens, ge(Low), Or).
expand_fuzzy(Token, Result) :-
    atomic(Token),
    !,
    (   rdf_stopgap_token(Token)
    ->  Result = @(stopgap)
    ;   Result = Token
    ).
expand_fuzzy(Token, _) :-
    throw(error(type_error(Token, boolean_expression), _)).

simplify(Expr0, Expr) :-
    simple(Expr0, Expr),
    !.
simplify(Expr, Expr).

simple(and(@(false), _), @(false)).
simple(and(_, @(false)), @(false)).
simple(and(@(stopgap), Token), Token).
simple(and(Token, @(stopgap)), Token).
simple(or(@(false), X), X).
simple(or(X, @(false)), X).
simple(or(@(stopgap), Token), Token).
simple(or(Token, @(stopgap)), Token).


list_to_or([], _, @(false)) :- !.
list_to_or([X], How, One) :-
    !,
    tag(How, X, One).
list_to_or([H0|T0], How, or(H, T)) :-
    tag(How, H0, H),
    list_to_or(T0, How, T).

tag(sounds(X),    Y, sounds(X,Y)).
tag(stem(X),      Y, stem(X,Y)).
tag(prefix(X),    Y, prefix(X,Y)).
tag(case(X),      Y, case(X,Y)).
tag(between(L,H), Y, between(L,H,Y)).
tag(ge(L),        Y, ge(L,Y)).
tag(le(H),        Y, le(H,Y)).

untag(sounds(_,Y),    Y).
untag(stem(_,Y),      Y).
untag(prefix(_,Y),    Y).
untag(case(_,Y),      Y).
untag(between(_,_,Y), Y).
untag(le(_,Y),        Y).
untag(ge(_,Y),        Y).

untag(sounds(X,Y),    sounds(X),    Y).
untag(stem(X,Y),      stem(X),      Y).
untag(prefix(X,Y),    prefix(X),    Y).
untag(case(X,Y),      case(X),      Y).
untag(between(L,H,Y), between(L,H), Y).
untag(ge(L,Y),        ge(L),        Y).
untag(le(H,Y),        le(H),        Y).


%!  nnf(+Formula, -NNF)
%
%   Rewrite to Negative Normal Form, meaning negations only appear
%   around literals.

nnf(not(not(A0)), A) :-
    !,
    nnf(A0, A).
nnf(not(and(A0,B0)), or(A,B)) :-
    !,
    nnf(not(A0), A),
    nnf(not(B0), B).
nnf(not(or(A0,B0)), and(A,B)) :-
    !,
    nnf(not(A0), A),
    nnf(not(B0), B).
nnf(A, A).


%!  dnf(+NNF, -DNF)
%
%   Convert a formula in NNF to Disjunctive Normal Form (DNF)

dnf(or(A0,B0), or(A, B)) :-
    !,
    dnf(A0, A),
    dnf(B0, B).
dnf(and(A0,B0), DNF):-
    !,
    dnf(A0, A1),
    dnf(B0, B1),
    dnf1(and(A1,B1), DNF).
dnf(DNF, DNF).

dnf1(and(A0, or(B,C)), or(P,Q)) :-
    !,
    dnf1(and(A0,B), P),
    dnf1(and(A0,C), Q).
dnf1(and(or(B,C), A0), or(P,Q)) :-
    !,
    dnf1(and(A0,B), P),
    dnf1(and(A0,C), Q).
dnf1(DNF, DNF).


                 /*******************************
                 *          TOKEN INDEX         *
                 *******************************/

%!  token_index(-Map)
%
%   Get the index of tokens. If  not   present,  create one from the
%   current database. Once created, the map is kept up-to-date using
%   a monitor hook.

token_index(Map) :-
    literal_map(token, Map),
    !,
    wait_for_map(token).
token_index(Map) :-
    rdf_new_literal_map(Map),
    assert(literal_map(token, Map)),
    register_token_updater,
    message_queue_create(Queue),
    assert(map_building(token, Queue)),
    thread_create(make_literal_index(Queue), _,
                  [ alias('__rdf_tokenizer'),
                    detached(true)
                  ]),
    wait_for_map(token).

register_token_updater :-
    Monitor = [ reset,
                new_literal,
                old_literal
              ],
    (   setting(index(default))
    ->  create_update_literal_thread(1),
        rdf_monitor(thread_monitor_literal, Monitor)
    ;   setting(index(thread(N)))
    ->  create_update_literal_thread(N),
        rdf_monitor(thread_monitor_literal, Monitor)
    ;   rdf_monitor(monitor_literal, Monitor)
    ).

make_literal_index(Queue) :-
    call_cleanup(
        make_literal_index,
        ( message_queue_destroy(Queue),
          retractall(map_building(token, _)))).

%!  make_literal_index
%
%   Create the initial literal index.

make_literal_index :-
    setting(index_threads(N)),
    !,
    threaded_literal_index(N),
    verbose('~N', []).
make_literal_index :-
    current_prolog_flag(cpu_count, X),
    threaded_literal_index(X),
    verbose('~N', []).

threaded_literal_index(N) :-
    N > 1,
    !,
    message_queue_create(Q, [max_size(1000)]),
    create_index_threads(N, Q, Ids),
    forall(rdf_current_literal(Literal),
           thread_send_message(Q, Literal)),
    forall(between(1, N, _),
           thread_send_message(Q, done(true))),
    maplist(thread_join, Ids, _).
threaded_literal_index(_) :-
    forall(rdf_current_literal(Literal),
           register_literal(Literal)).

create_index_threads(N, Q, [Id|T]) :-
    N > 0,
    !,
    thread_create(index_worker(Q), Id, []),
    N2 is N - 1,
    create_index_threads(N2, Q, T).
create_index_threads(_, _, []) :- !.

index_worker(Queue) :-
    repeat,
        thread_get_message(Queue, Msg),
        work(Msg).

work(done(true)) :- !.
work(Literal) :-
    register_literal(Literal),
    fail.


%!  clean_token_index
%
%   Clean after a reset.

clean_token_index :-
    forall(literal_map(_, Map),
           rdf_reset_literal_map(Map)),
    retractall(stopgap(_)).

%!  rdf_delete_literal_index(+Type)
%
%   Fully delete a literal index

rdf_delete_literal_index(Type) :-
    must_be(atom, Type),
    (   retract(literal_map(Type, Map))
    ->  rdf_reset_literal_map(Map)          % destroy is unsafe
    ).

                 /*******************************
                 *        THREADED UPDATE       *
                 *******************************/

%!  create_update_literal_thread(+Threads)
%
%   Setup literal monitoring using threads.  While loading databases
%   through rdf_attach_db/2 from  rdf_persistency.pl,   most  of the
%   time is spent updating the literal token database. While loading
%   the RDF triples, most of the time   is spend in updating the AVL
%   tree holding the literals. Updating  the   token  index hangs on
%   updating the AVL trees holding the   tokens.  Both tasks however
%   can run concurrently.

create_update_literal_thread(Threads) :-
    message_queue_create(_,
                         [ alias(rdf_literal_monitor_queue),
                           max_size(50000)
                         ]),
    forall(between(1, Threads, _),
           create_index_worker(initial)).

:- dynamic
    index_worker_id/1,
    extra_worker_count/1.

create_index_worker(Status) :-
    (   retract(index_worker_id(Id0))
    ->  true
    ;   Id0 = 1
    ),
    succ(Id0, Id1),
    assertz(index_worker_id(Id1)),
    atom_concat(rdf_literal_monitor_, Id0, Alias),
    inc_extra_worker_count(Status),
    thread_create(monitor_literals(Status), _,
                  [ alias(Alias)
                  ]).

monitor_literals(initial) :-
    set_prolog_flag(agc_margin, 0), % we don't create garbage
    repeat,
        thread_get_message(rdf_literal_monitor_queue, Literal),
        register_literal(Literal),
    fail.
monitor_literals(extra) :-
    set_prolog_flag(agc_margin, 0),
    repeat,
        (   thread_get_message(rdf_literal_monitor_queue, Literal,
                               [ timeout(1)
                               ])
        ->  register_literal(Literal),
            fail
        ;   !
        ),
    with_mutex(create_index_worker, dec_extra_worker_count),
    thread_self(Me),
    thread_detach(Me).

thread_monitor_literal(new_literal(Literal)) :-
    !,
    thread_send_message(rdf_literal_monitor_queue, Literal).
thread_monitor_literal(Action) :-
    !,
    monitor_literal(Action).

%!  check_index_workers(+Queue, +Keys)
%
%   Increase the number of workers indexing   literals sent to Queue
%   if the queue gets overfull.

check_index_workers(Alias, Keys) :-
    max_extra_workers(Max),
    Max > 0,
    message_queue_property(Queue, alias(Alias)),
    message_queue_property(Queue, size(Size)),
    Size > 10000,
    \+ ( extra_worker_count(Extra),
         Extra >= Max
       ),
    !,
    debug(rdf_litindex,
          'Creating extra literal indexer (Queue=~D, Keys=~D)',
          [Size, Keys]),
    with_mutex(create_index_worker, create_index_worker(extra)).
check_index_workers(_, _).

inc_extra_worker_count(extra) :-
    !,
    (   retract(extra_worker_count(C0))
    ->  C is C0+1
    ;   C = 1
    ),
    asserta(extra_worker_count(C)).
inc_extra_worker_count(_).

dec_extra_worker_count :-
    retract(extra_worker_count(C0)),
    !,
    C is C0-1,
    asserta(extra_worker_count(C)).
dec_extra_worker_count.

max_extra_workers(Max) :-
    current_prolog_flag(cpu_count, Count),
    Max is Count//2.


                 /*******************************
                 *       MONITORED UPDATE       *
                 *******************************/

monitor_literal(new_literal(Literal)) :-
    register_literal(Literal).
monitor_literal(old_literal(Literal)) :-
    unregister_literal(Literal).
monitor_literal(transaction(begin, reset)) :-
    rdf_monitor(monitor_literal, [-old_literal]),
    clean_token_index.
monitor_literal(transaction(end, reset)) :-
    rdf_monitor(monitor_literal, [+old_literal]).

%!  register_literal(+Literal)
%
%   Associate the tokens of a literal with the literal itself.

register_literal(Literal) :-
    (   rdf_tokenize_literal(Literal, Tokens0)
    ->  sort(Tokens0, Tokens),
        text_of(Literal, Lang, Text),
        literal_map(token, Map),
        add_tokens(Tokens, Lang, Text, Map)
    ;   true
    ).

add_tokens([], _, _, _).
add_tokens([H|T], Lang, Literal, Map) :-
    rdf_insert_literal_map(Map, H, Literal, Keys),
    (   var(Keys)
    ->  (   rdf_keys_in_literal_map(Map, key(H), Count),
            setting(stopgap_threshold(Threshold)),
            Count > Threshold
        ->  assert(stopgap(H)),
            rdf_delete_literal_map(Map, H)
        ;   true
        )
    ;   forall(new_token(H, Lang), true),
        (   Keys mod 1000 =:= 0
        ->  progress(Map, 'Tokens'),
            (   Keys mod 10000 =:= 0
            ->  check_index_workers(rdf_literal_monitor_queue, Keys)
            ;   true
            )
        ;   true
        )
    ),
    add_tokens(T, Lang, Literal, Map).


%!  unregister_literal(+Literal)
%
%   Literal is removed from the database.   As we abstract from lang
%   and type qualifiers we first have to  check this is the last one
%   that is destroyed.

unregister_literal(Literal) :-
    text_of(Literal, _Lang, Text),
    (   rdf(_,_,literal(Text))
    ->  true                        % still something left
    ;   rdf_tokenize_literal(Literal, Tokens0),
        sort(Tokens0, Tokens),
        literal_map(token, Map),
        del_tokens(Tokens, Text, Map)
    ).

del_tokens([], _, _).
del_tokens([H|T], Literal, Map) :-
    rdf_delete_literal_map(Map, H, Literal),
    del_tokens(T, Literal, Map).


%!  rdf_tokenize_literal(+Literal, -Tokens) is semidet.
%
%   Tokenize a literal. We make  this   hookable  as tokenization is
%   generally domain dependent.

rdf_tokenize_literal(Literal, Tokens) :-
    tokenization(Literal, Tokens),
    !.               % Hook
rdf_tokenize_literal(Literal, Tokens) :-
    text_of(Literal, _Lang, Text),
    atom(Text),
    tokenize_atom(Text, Tokens0),
    select_tokens(Tokens0, Tokens).

select_tokens([], []).
select_tokens([H|T0], T) :-
    (   exclude_from_index(token, H)
    ->  select_tokens(T0, T)
    ;   number(H)
    ->  (   integer(H),
            between(-1073741824, 1073741823, H)
        ->  T = [H|T1],
            select_tokens(T0, T1)
        ;   select_tokens(T0, T)
        )
    ;   atom_length(H, 1)
    ->  select_tokens(T0, T)
    ;   default_stopgap(H)
    ->  select_tokens(T0, T)
    ;   stopgap(H)
    ->  select_tokens(T0, T)
    ;   T = [H|T1],
        select_tokens(T0, T1)
    ).

%!  rdf_stopgap_token(-Token) is nondet.
%
%   True when Token is a stopgap  token. Currently, this implies one
%   of:
%
%     - exclude_from_index(token, Token) is true
%     - default_stopgap(Token) is true
%     - Token is an atom of length 1
%     - Token was added to the dynamic stopgap token set because
%       it appeared in more than _stopgap_threshold_ literals.

rdf_stopgap_token(Token) :-
    (   var(Token)
    ->  rdf_stopgap_token2(Token)
    ;   rdf_stopgap_token2(Token), !
    ).

rdf_stopgap_token2(Token) :-
    exclude_from_index(token, Token).
rdf_stopgap_token2(Token) :-
    default_stopgap(Token).
rdf_stopgap_token2(Token) :-
    atom(Token),
    atom_length(Token, 1).
rdf_stopgap_token2(Token) :-
    stopgap(Token).

%!  default_stopgap(?Token)
%
%   Tokens we do not wish to index,  as  they create huge amounts of
%   data with little or no value.  Is   there  a more general way to
%   describe this? Experience shows that simply  word count is not a
%   good criterion as it often rules out popular domain terms.

default_stopgap(and).
default_stopgap(an).
default_stopgap(or).
default_stopgap(of).
default_stopgap(on).
default_stopgap(in).
default_stopgap(this).
default_stopgap(the).


%!  text_of(+LiteralArg, -Lang, -Text) is semidet.
%
%   Get the textual  or  (integer)   numerical  information  from  a
%   literal value. Lang  is  the  language   to  use  for  stemming.
%   Currently we use English for untyped  plain literals or literals
%   typed xsd:string. Formally, these should not be tokenized, but a
%   lot of data out there does not tag strings with their language.

text_of(type(xsd:string, Text), en, Text) :- !.
text_of(type(_, Text), -, Text) :- !.
text_of(lang(Lang, Text), Lang, Text) :- !.
text_of(Text, en, Text) :- atom(Text), !.
text_of(Text, -, Text) :- integer(Text).


                 /*******************************
                 *         STEM INDEX           *
                 *******************************/

%!  stem_index(-Map) is det.
%
%   Get the stemming literal index. This index is created on demand.
%   If some thread is creating the index, other threads wait for its
%   completion.

stem_index(Map) :-
    literal_map(stem, Map),
    !,
    wait_for_map(stem).
stem_index(Map) :-
    rdf_new_literal_map(Map),
    assert(literal_map(stem, Map)),
    assert((new_token(Token, Lang) :- add_stem(Token, Lang, Map))),
    message_queue_create(Queue),
    assert(map_building(stem, Queue)),
    thread_create(fill_stem_index(Map, Queue), _,
                  [ alias('__rdf_stemmer'),
                    detached(true)
                  ]),
    wait_for_map(stem).

wait_for_map(MapName) :-
    (   map_building(MapName, Queue)
    ->  catch(thread_get_message(Queue, _), _, true),
        wait_for_map(MapName)
    ;   true
    ).

fill_stem_index(StemMap, Queue) :-
    call_cleanup(
        forall(rdf_current_literal(Literal),
               stem_literal_tokens(Literal, StemMap)),
        ( message_queue_destroy(Queue),
          retractall(map_building(stem, _)))).

stem_literal_tokens(Literal, StemMap) :-
    rdf_tokenize_literal(Literal, Tokens),
    !,
    sort(Tokens, Tokens1),
    text_of(Literal, Lang, _Text),
    insert_tokens_stem(Tokens1, Lang, StemMap).
stem_literal_tokens(_,_).

insert_tokens_stem([], _, _).
insert_tokens_stem([Token|T], Lang, Map) :-
    (   atom(Token)
    ->  (   stem(Token, Lang, Stem)
        ->  rdf_insert_literal_map(Map, Stem, Token, Keys),
            (   integer(Keys),
                Keys mod 1000 =:= 0
            ->  progress(Map, 'Stem')
            ;   true
            )
        ;   true
        )
    ;   true
    ),
    insert_tokens_stem(T, Lang, Map).


add_stem(Token, Lang, Map) :-
    stem(Lang, Token, Stem),
    rdf_insert_literal_map(Map, Stem, Token, _).

stem(Token, LangSpec, Stem) :-
    main_lang(LangSpec, Lang),
    downcase_atom(Token, Lower),
    catch(snowball(Lang, Lower, Stem), _, fail).

main_lang(LangSpec, Lang) :-
    sub_atom(LangSpec, Before, _, _, -),
    !,
    sub_atom(LangSpec, 0, Before, _, Lang).
main_lang(LangSpec, Lang) :-
    downcase_atom(LangSpec, Lang).


                 /*******************************
                 *        METAPHONE INDEX       *
                 *******************************/


metaphone_index(Map) :-
    literal_map(metaphone, Map),
    !,
    wait_for_map(metaphone).
metaphone_index(Map) :-
    rdf_new_literal_map(Map),
    assert(literal_map(metaphone, Map)),
    assert((new_token(Token, Lang) :- add_metaphone(Token, Lang, Map))),
    message_queue_create(Queue),
    assert(map_building(metaphone, Queue)),
    thread_create(fill_metaphone_index(Map, Queue), _,
                  [ alias('__rdf_metaphone_indexer'),
                    detached(true)
                  ]),
    wait_for_map(metaphone).

fill_metaphone_index(MetaphoneMap, Queue) :-
    call_cleanup(
        fill_metaphone_index(MetaphoneMap),
        ( message_queue_destroy(Queue),
          retractall(map_building(metaphone, _)))).

fill_metaphone_index(MetaphoneMap) :-
    token_index(TokenMap),
    rdf_keys_in_literal_map(TokenMap, all, Tokens),
    metaphone(Tokens, MetaphoneMap).

metaphone([], _).
metaphone([Token|T], Map) :-
    (   atom(Token),
        double_metaphone(Token, SoundEx)
    ->  rdf_insert_literal_map(Map, SoundEx, Token, Keys),
        (   integer(Keys),
            Keys mod 1000 =:= 0
        ->  progress(Map, 'Metaphone')
        ;   true
        )
    ;   true
    ),
    metaphone(T, Map).


add_metaphone(Token, _Lang, Map) :-
    atom(Token),
    !,
    double_metaphone(Token, SoundEx),
    rdf_insert_literal_map(Map, SoundEx, Token).
add_metaphone(_, _, _).

%!  rdf_literal_index(+Type, -Index) is det.
%
%   True when Index is a literal map   containing the index of Type.
%   Type is one of:
%
%     - token
%     Tokens are basically words of literal values. See
%     rdf_tokenize_literal/2.  The `token` map maps tokens to full
%     literal texts.
%     - stem
%     Index of stemmed tokens.  If the language is available, the
%     tokens are stemmed using the matching _snowball_ stemmer.
%     The `stem` map maps stemmed to full tokens.
%     - metaphone
%     Phonetic index of tokens.  The `metaphone` map maps phonetic
%     keys to tokens.

rdf_literal_index(token, Map) :-
    !,
    token_index(Map).
rdf_literal_index(stem, Map) :-
    !,
    stem_index(Map).
rdf_literal_index(metaphone, Map) :-
    !,
    metaphone_index(Map).
rdf_literal_index(Type, _Map) :-
    domain_error(literal_index, Type).


                 /*******************************
                 *             UTIL             *
                 *******************************/

verbose(Fmt, Args) :-
    setting(verbose(true)),
    !,
    format(user_error, Fmt, Args).
verbose(_, _).

progress(Map, Which) :-
    setting(verbose(true)),
    !,
    rdf_statistics_literal_map(Map, size(Keys, Values)),
    format(user_error,
           '\r~t~w: ~12|Keys: ~t~D~15+; Values: ~t~D~20+',
           [Which, Keys, Values]).
progress(_,_).