msyds/gf-core
1
0
Fork 0
forked from GitHub/gf-core
Code Pull Requests Activity

the content of ParseEngAbs3.probs is now merged with ParseEngAbs.probs. The later is now retrained. Once the grammar is compiled with the .probs file now it doesn't need anything more to do robust parsing. The robustness itself is controlled by the flags 'heuristic_search_factor', 'meta_prob' and 'meta_token_prob' in ParseEngAbs.gf

Browse Source
This commit is contained in:
kr.angelov
2013-11-06 10:21:46 +00:00
parent 84ef5fa5fa
commit 2483dc7728
27 changed files with 65052 additions and 65106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/compiler/GF/Command/Commands.hs
View File

@@ -1134,15 +1134,19 @@ allCommands = Map.fromList [
case arg of
[EFun id] -> case Map.lookup id (funs (abstract pgf)) of
Just fd -> do putStrLn $ render (ppFun id fd)
putStrLn ("Probability: "++show (probTree pgf (EFun id)))
let (_,_,_,prob,_) = fd
putStrLn ("Probability: "++show prob)
return void
Nothing -> case Map.lookup id (cats (abstract pgf)) of
Just hyps -> do putStrLn $
render (ppCat id hyps $$
Just cd -> do putStrLn $
render (ppCat id cd $$
if null (functionsToCat pgf id)
then empty
else space $$
vcat [ppFun fid (ty,0,Just [],0,0) | (fid,ty) <- functionsToCat pgf id])
vcat [ppFun fid (ty,0,Just [],0,0) | (fid,ty) <- functionsToCat pgf id] $$
space)
let (_,_,prob,_) = cd
putStrLn ("Probability: "++show prob)
return void
Nothing -> do putStrLn ("unknown category of function identifier "++show id)
return void

6
src/compiler/GF/Compile/GrammarToPGF.hs
View File

@@ -50,12 +50,12 @@ mkCanon2pgf opts gr am = do
[((cPredefAbs,c), AbsCat (Just (L NoLoc []))) | c <- [cFloat,cInt,cString]] ++
Look.allOrigInfos gr am
flags = Map.fromList [(mkCId f,if f == "beam_size" then C.LFlt (read x) else C.LStr x) | (f,x) <- optionsPGF aflags]
flags = Map.fromList [(mkCId f,x) | (f,x) <- optionsPGF aflags]
funs = Map.fromList [(i2i f, (mkType [] ty, mkArrity ma, mkDef pty, 0, addr)) |
((m,f),AbsFun (Just (L _ ty)) ma pty _,addr) <- adefs]
cats = Map.fromList [(i2i c, (snd (mkContext [] cont),catfuns c, addr)) |
cats = Map.fromList [(i2i c, (snd (mkContext [] cont),catfuns c, 0, addr)) |
((m,c),AbsCat (Just (L _ cont)),addr) <- adefs]
catfuns cat =
@@ -69,7 +69,7 @@ mkCanon2pgf opts gr am = do
([((cPredefAbs,c), CncCat (Just (L NoLoc GM.defLinType)) Nothing Nothing Nothing Nothing) | c <- [cInt,cFloat,cString]] ++
Look.allOrigInfos gr cm)
let flags = Map.fromList [(mkCId f,if f == "beam_size" then C.LFlt (read x) else C.LStr x) | (f,x) <- optionsPGF cflags]
let flags = Map.fromList [(mkCId f,x) | (f,x) <- optionsPGF cflags]
seqs = (mkSetArray . Set.fromList . concat) $
(Map.keys ex_seqs : [maybe [] elems (mseqs mi) | (m,mi) <- allExtends gr cm])

6
src/compiler/GF/Compile/PGFtoLProlog.hs
View File

@@ -12,7 +12,7 @@ import qualified Data.Map as Map
grammar2lambdaprolog_mod pgf = render $
text "module" <+> ppCId (absname pgf) <> char '.' $$
space $$
vcat [ppClauses cat fns | (cat,(_,fs,_)) <- Map.toList (cats (abstract pgf)),
vcat [ppClauses cat fns | (cat,(_,fs,_,_)) <- Map.toList (cats (abstract pgf)),
let fns = [(f,fromJust (Map.lookup f (funs (abstract pgf)))) | (_,f) <- fs]]
where
ppClauses cat fns =
@@ -25,11 +25,11 @@ grammar2lambdaprolog_mod pgf = render $
grammar2lambdaprolog_sig pgf = render $
text "sig" <+> ppCId (absname pgf) <> char '.' $$
space $$
vcat [ppCat c hyps <> dot | (c,(hyps,_,_)) <- Map.toList (cats (abstract pgf))] $$
vcat [ppCat c hyps <> dot | (c,(hyps,_,_,_)) <- Map.toList (cats (abstract pgf))] $$
space $$
vcat [ppFun f ty <> dot | (f,(ty,_,Nothing,_,_)) <- Map.toList (funs (abstract pgf))] $$
space $$
vcat [ppExport c hyps <> dot | (c,(hyps,_,_)) <- Map.toList (cats (abstract pgf))] $$
vcat [ppExport c hyps <> dot | (c,(hyps,_,_,_)) <- Map.toList (cats (abstract pgf))] $$
vcat [ppFunPred f (hyps ++ [(Explicit,wildCId,DTyp [] c es)]) <> dot | (f,(DTyp hyps c es,_,Just _,_,_)) <- Map.toList (funs (abstract pgf))]
ppCat :: CId -> [Hypo] -> Doc

2
src/compiler/GF/Compile/PGFtoProlog.hs
View File

@@ -49,7 +49,7 @@ plAbstract name abs
(f, v) <- Map.assocs (aflags abs)] ++++
plFacts name "cat" 2 "(?Type, ?[X:Type,...])"
[[plType cat args, plHypos hypos'] |
(cat, (hypos, _, _)) <- Map.assocs (cats abs),
(cat, (hypos,_,_,_)) <- Map.assocs (cats abs),
let ((_, subst), hypos') = mapAccumL alphaConvertHypo emptyEnv hypos,
let args = reverse [EFun x | (_,x) <- subst]] ++++
plFacts name "fun" 3 "(?Fun, ?Type, ?[X:Type,...])"

6
src/compiler/GF/Grammar/Binary.hs
View File

@@ -17,6 +17,7 @@ import Data.Binary
--import Control.Monad
import qualified Data.Map as Map
import qualified Data.ByteString.Char8 as BS
import Text.PrettyPrint (render)
import GF.Data.Operations
import GF.Infra.Ident
@@ -24,9 +25,10 @@ import GF.Infra.Option
import GF.Grammar.Grammar
import PGF() -- Binary instances
import PGF.Data(ppLit)
-- Please change this every time when the GFO format is changed
gfoVersion = "GF02"
gfoVersion = "GF03"
instance Binary Ident where
put id = put (ident2bs id)
@@ -91,7 +93,7 @@ instance Binary ModuleStatus where
instance Binary Options where
put = put . optionsGFO
get = do opts <- get
case parseModuleOptions ["--" ++ flag ++ "=" ++ value | (flag,value) <- opts] of
case parseModuleOptions ["--" ++ flag ++ "=" ++ render (ppLit value) | (flag,value) <- opts] of
Ok x -> return x
Bad msg -> fail msg

4
src/compiler/GF/Grammar/Printer.hs
View File

@@ -29,7 +29,7 @@ import GF.Infra.Option
import GF.Grammar.Values
import GF.Grammar.Grammar
import PGF.Data (ppMeta)
import PGF.Data (ppMeta, ppLit)
import PGF.Printer (ppFId, ppFunId, ppSeqId, ppSeq)
import Text.PrettyPrint
@@ -87,7 +87,7 @@ ppModule q (mn, ModInfo mtype mstat opts exts with opens _ _ mseqs jments) =
ppOptions opts =
text "flags" $$
nest 2 (vcat [text option <+> equals <+> str value <+> semi | (option,value) <- optionsGFO opts])
nest 2 (vcat [text option <+> equals <+> ppLit value <+> semi | (option,value) <- optionsGFO opts])
ppJudgement q (id, AbsCat pcont ) =
text "cat" <+> ppIdent id <+>

33
src/compiler/GF/Infra/Option.hs
View File

@@ -43,6 +43,7 @@ import GF.Data.ErrM
import Data.Set (Set)
import qualified Data.Set as Set
import PGF.Data(Literal(..))
usageHeader :: String
usageHeader = unlines
@@ -170,7 +171,9 @@ data Flags = Flags {
optWarnings :: [Warning],
optDump :: [Dump],
optTagsOnly :: Bool,
optBeamSize :: Maybe Double,
optHeuristicFactor :: Maybe Double,
optMetaProb :: Maybe Double,
optMetaToknProb :: Maybe Double,
optNewComp :: Bool
}
deriving (Show)
@@ -206,16 +209,18 @@ fixRelativeLibPaths curr_dir lib_dir (Options o) = Options (fixPathFlags . o)
-- Showing options
-- | Pretty-print the options that are preserved in .gfo files.
optionsGFO :: Options -> [(String,String)]
optionsGFO :: Options -> [(String,Literal)]
optionsGFO opts = optionsPGF opts
++ [("coding", flag optEncoding opts)]
++ [("coding", LStr (flag optEncoding opts))]
-- | Pretty-print the options that are preserved in .pgf files.
optionsPGF :: Options -> [(String,String)]
optionsPGF :: Options -> [(String,Literal)]
optionsPGF opts =
maybe [] (\x -> [("language",x)]) (flag optSpeechLanguage opts)
++ maybe [] (\x -> [("startcat",x)]) (flag optStartCat opts)
++ maybe [] (\x -> [("beam_size",show x)]) (flag optBeamSize opts)
maybe [] (\x -> [("language",LStr x)]) (flag optSpeechLanguage opts)
++ maybe [] (\x -> [("startcat",LStr x)]) (flag optStartCat opts)
++ maybe [] (\x -> [("heuristic_search_factor",LFlt x)]) (flag optHeuristicFactor opts)
++ maybe [] (\x -> [("meta_prob",LFlt x)]) (flag optMetaProb opts)
++ maybe [] (\x -> [("meta_token_prob",LFlt x)]) (flag optMetaToknProb opts)
-- Option manipulation
@@ -272,7 +277,9 @@ defaultFlags = Flags {
optWarnings = [],
optDump = [],
optTagsOnly = False,
optBeamSize = Nothing,
optHeuristicFactor = Nothing,
optMetaProb = Nothing,
optMetaToknProb = Nothing,
optNewComp =
#ifdef NEW_COMP
True
@@ -358,7 +365,9 @@ optDescr =
Option [] ["stem"] (onOff (toggleOptimize OptStem) True) "Perform stem-suffix analysis (default on).",
Option [] ["cse"] (onOff (toggleOptimize OptCSE) True) "Perform common sub-expression elimination (default on).",
Option [] ["cfg"] (ReqArg cfgTransform "TRANS") "Enable or disable specific CFG transformations. TRANS = merge, no-merge, bottomup, no-bottomup, ...",
Option [] ["beam_size"] (ReqArg readDouble "SIZE") "Set the beam size for statistical parsing",
Option [] ["heuristic_search_factor"] (ReqArg (readDouble (\d o -> o { optHeuristicFactor = Just d })) "FACTOR") "Set the heuristic search factor for statistical parsing",
Option [] ["meta_prob"] (ReqArg (readDouble (\d o -> o { optMetaProb = Just d })) "PROB") "Set the probability of introducting a meta variable in the parser",
Option [] ["meta_token_prob"] (ReqArg (readDouble (\d o -> o { optMetaToknProb = Just d })) "PROB") "Set the probability for skipping a token in the parser",
Option [] ["new-comp"] (NoArg (set $ \o -> o{optNewComp = True})) "Use the new experimental compiler.",
Option [] ["old-comp"] (NoArg (set $ \o -> o{optNewComp = False})) "Use old trusty compiler.",
dumpOption "source" Source,
@@ -433,9 +442,9 @@ optDescr =
Nothing -> fail $ "Unknown CFG transformation: " ++ x'
++ " Known: " ++ show (map fst cfgTransformNames)
readDouble x = case reads x of
[(d,"")] -> set $ \o -> o { optBeamSize = Just d }
_ -> fail "A floating point number is expected"
readDouble f x = case reads x of
[(d,"")] -> set $ f d
_ -> fail "A floating point number is expected"
dumpOption s d = Option [] ["dump-"++s] (NoArg (set $ \o -> o { optDump = Dump d:optDump o})) ("Dump output of the " ++ s ++ " phase.")

2
src/compiler/GF/Speech/VoiceXML.hs
View File

@@ -39,7 +39,7 @@ type Skeleton = [(CId, [(CId, [CId])])]
pgfSkeleton :: PGF -> Skeleton
pgfSkeleton pgf = [(c,[(f,fst (catSkeleton (lookType (abstract pgf) f))) | (_,f) <- fs])
| (c,(_,fs,_)) <- Map.toList (cats (abstract pgf))]
| (c,(_,fs,_,_)) <- Map.toList (cats (abstract pgf))]
--
-- * Questions to ask

2
src/compiler/GFC.hs
View File

@@ -104,7 +104,7 @@ writeByteCode opts pgf
where
addrs =
[(id,addr) | (id,(_,_,_,_,addr)) <- Map.toList (funs (abstract pgf))] ++
[(id,addr) | (id,(_,_,addr)) <- Map.toList (cats (abstract pgf))]
[(id,addr) | (id,(_,_,_,addr)) <- Map.toList (cats (abstract pgf))]
writePGF :: Options -> PGF -> IOE ()
writePGF opts pgf = do

6
src/compiler/SimpleEditor/Convert.hs
View File

@@ -23,6 +23,7 @@ import GF.Grammar.Printer(ppParams,ppTerm,getAbs,TermPrintQual(..))
import GF.Grammar.Parser(runP,pModDef)
import GF.Grammar.Lexer(Posn(..))
import GF.Data.ErrM
import PGF.Data(Literal(LStr))
import SimpleEditor.Syntax as S
import SimpleEditor.JSON
@@ -57,7 +58,10 @@ convAbstract (modid,src) =
let cats = reverse cats0
funs = reverse funs0
flags = optionsGFO (mflags src)
startcat = maybe "-" id $ lookup "startcat" flags
startcat =
case lookup "startcat" flags of
Just (LStr cat) -> cat
_ -> "-"
return $ Grammar (convId modid) extends (Abstract startcat cats funs) []
convExtends = mapM convExtend

5
src/runtime/c/pgf/data.h
View File

@@ -87,9 +87,7 @@ typedef struct {
PgfCId name;
PgfHypos* context;
prob_t meta_prob;
prob_t meta_token_prob;
PgfMetaChildMap* meta_child_probs;
prob_t prob;
void* predicate;
} PgfAbsCat;
@@ -230,6 +228,7 @@ typedef GuSeq PgfCncFuns;
struct PgfConcr {
PgfCId name;
PgfAbstr* abstr;
PgfFlags* cflags;
PgfPrintNames* printnames;
GuMap* ccats;

92
src/runtime/c/pgf/parser.c
View File

@@ -63,7 +63,10 @@ typedef struct {
int prod_full_count;
#endif
PgfItem* free_item;
prob_t beam_size;
prob_t heuristic_factor;
prob_t meta_prob;
prob_t meta_token_prob;
} PgfParsing;
typedef enum { BIND_NONE, BIND_HARD, BIND_SOFT } BIND_TYPE;
@@ -1389,12 +1392,14 @@ pgf_parsing_meta_predict(GuMapItor* fn, const void* key, void* value, GuExn* err
{
(void) (err);
PgfAbsCat* abscat = (PgfAbsCat*) key;
prob_t meta_prob = *((prob_t*) value);
PgfAbsCat* abscat = *((PgfAbsCat**) value);
PgfMetaPredictFn* clo = (PgfMetaPredictFn*) fn;
PgfParsing* ps = clo->ps;
PgfItem* meta_item = clo->meta_item;
if (abscat->prob == INFINITY)
return;
PgfCncCat* cnccat =
gu_map_get(ps->concr->cnccats, abscat->name, PgfCncCat*);
if (cnccat == NULL)
@@ -1412,7 +1417,7 @@ pgf_parsing_meta_predict(GuMapItor* fn, const void* key, void* value, GuExn* err
PgfItem* item =
pgf_item_copy(meta_item, ps);
item->inside_prob +=
ccat->viterbi_prob+meta_prob;
ccat->viterbi_prob+abscat->prob;
size_t nargs = gu_seq_length(meta_item->args);
item->args = gu_new_seq(PgfPArg, nargs+1, ps->pool);
@@ -1698,18 +1703,14 @@ pgf_parsing_item(PgfParsing* ps, PgfItem* item)
}
pgf_parsing_complete(ps, item, ep);
} else {
prob_t meta_token_prob =
item->conts->ccat->cnccat->abscat->meta_token_prob;
prob_t meta_token_prob =
ps->meta_token_prob;
if (meta_token_prob != INFINITY) {
pgf_parsing_meta_scan(ps, item, meta_token_prob);
}
PgfCIdMap* meta_child_probs =
item->conts->ccat->cnccat->abscat->meta_child_probs;
if (meta_child_probs != NULL) {
PgfMetaPredictFn clo = { { pgf_parsing_meta_predict }, ps, item };
gu_map_iter(meta_child_probs, &clo.fn, NULL);
}
PgfMetaPredictFn clo = { { pgf_parsing_meta_predict }, ps, item };
gu_map_iter(ps->concr->abstr->cats, &clo.fn, NULL);
}
} else {
pgf_parsing_symbol(ps, item, item->curr_sym);
@@ -1721,22 +1722,38 @@ pgf_parsing_item(PgfParsing* ps, PgfItem* item)
}
}
static prob_t
pgf_parsing_default_beam_size(PgfConcr* concr)
static void
pgf_parsing_set_default_factors(PgfParsing* ps, PgfAbstr* abstr)
{
PgfLiteral lit = gu_map_get(concr->cflags, "beam_size", PgfLiteral);
PgfLiteral lit;
if (gu_variant_is_null(lit))
return 0;
lit =
gu_map_get(abstr->aflags, "heuristic_search_factor", PgfLiteral);
if (!gu_variant_is_null(lit)) {
GuVariantInfo pi = gu_variant_open(lit);
gu_assert (pi.tag == PGF_LITERAL_FLT);
ps->heuristic_factor = ((PgfLiteralFlt*) pi.data)->val;
}
GuVariantInfo pi = gu_variant_open(lit);
gu_assert (pi.tag == PGF_LITERAL_FLT);
return ((PgfLiteralFlt*) pi.data)->val;
lit =
gu_map_get(abstr->aflags, "meta_prob", PgfLiteral);
if (!gu_variant_is_null(lit)) {
GuVariantInfo pi = gu_variant_open(lit);
gu_assert (pi.tag == PGF_LITERAL_FLT);
ps->meta_prob = - log(((PgfLiteralFlt*) pi.data)->val);
}
lit =
gu_map_get(abstr->aflags, "meta_token_prob", PgfLiteral);
if (!gu_variant_is_null(lit)) {
GuVariantInfo pi = gu_variant_open(lit);
gu_assert (pi.tag == PGF_LITERAL_FLT);
ps->meta_token_prob = - log(((PgfLiteralFlt*) pi.data)->val);
}
}
static PgfParsing*
pgf_new_parsing(PgfConcr* concr,
GuString sentence, double heuristics,
pgf_new_parsing(PgfConcr* concr, GuString sentence,
GuPool* pool, GuPool* out_pool)
{
PgfParsing* ps = gu_new(PgfParsing, pool);
@@ -1756,7 +1773,11 @@ pgf_new_parsing(PgfConcr* concr,
ps->prod_full_count = 0;
#endif
ps->free_item = NULL;
ps->beam_size = heuristics;
ps->heuristic_factor = 0;
ps->meta_prob = INFINITY;
ps->meta_token_prob = INFINITY;
pgf_parsing_set_default_factors(ps, concr->abstr);
PgfExprMeta *expr_meta =
gu_new_variant(PGF_EXPR_META,
@@ -2107,7 +2128,7 @@ pgf_parse_result_is_new(PgfExprState* st)
// TODO: s/CId/Cat, add the cid to Cat, make Cat the key to CncCat
static PgfParsing*
pgf_parsing_init(PgfConcr* concr, PgfCId cat, size_t lin_idx,
GuString sentence, double heuristics,
GuString sentence, double heuristic_factor,
GuExn* err,
GuPool* pool, GuPool* out_pool)
{
@@ -2121,12 +2142,13 @@ pgf_parsing_init(PgfConcr* concr, PgfCId cat, size_t lin_idx,
gu_assert(lin_idx < cnccat->n_lins);
if (heuristics < 0) {
heuristics = pgf_parsing_default_beam_size(concr);
PgfParsing* ps =
pgf_new_parsing(concr, sentence, pool, out_pool);
if (heuristic_factor >= 0) {
ps->heuristic_factor = heuristic_factor;
}
PgfParsing* ps =
pgf_new_parsing(concr, sentence, heuristics, pool, out_pool);
PgfParseState* state =
pgf_new_parse_state(ps, 0, BIND_SOFT);
@@ -2156,11 +2178,13 @@ pgf_parsing_init(PgfConcr* concr, PgfCId cat, size_t lin_idx,
gu_buf_heap_push(state->agenda, pgf_item_prob_order, &item);
}
PgfItem *item =
pgf_new_item(ps, conts, ps->meta_prod);
item->inside_prob =
ccat->cnccat->abscat->meta_prob;
gu_buf_heap_push(state->agenda, pgf_item_prob_order, &item);
if (ps->meta_prob != INFINITY) {
PgfItem *item =
pgf_new_item(ps, conts, ps->meta_prod);
item->inside_prob =
ps->meta_prob;
gu_buf_heap_push(state->agenda, pgf_item_prob_order, &item);
}
}
}
@@ -2200,7 +2224,7 @@ pgf_parsing_proceed(PgfParsing* ps)
prob_t state_delta =
(st->viterbi_prob-(st->next ? st->next->viterbi_prob : 0))*
ps->beam_size;
ps->heuristic_factor;
delta_prob += state_delta;
st = st->next;
}

57
src/runtime/c/pgf/pgf.c
View File

@@ -35,63 +35,6 @@ pgf_read(const char* fpath,
return pgf;
}
void
pgf_load_meta_child_probs(PgfPGF* pgf, const char* fpath,
GuPool* pool, GuExn* err)
{
FILE *fp = fopen(fpath, "r");
if (!fp) {
gu_raise_errno(err);
return;
}
GuPool* tmp_pool = gu_new_pool();
for (;;) {
char cat1[21];
char cat2[21];
prob_t prob;
if (fscanf(fp, "%20s\t%20s\t%f", cat1, cat2, &prob) < 3)
break;
prob = - log(prob);
PgfAbsCat* abscat1 =
gu_map_get(pgf->abstract.cats, cat1, PgfAbsCat*);
if (abscat1 == NULL) {
GuExnData* exn = gu_raise(err, PgfExn);
exn->data = "Unknown category name";
goto close;
}
if (strcmp(cat2, "*") == 0) {
abscat1->meta_prob = prob;
} else if (strcmp(cat2, "_") == 0) {
abscat1->meta_token_prob = prob;
} else {
PgfAbsCat* abscat2 = gu_map_get(pgf->abstract.cats, cat2, PgfAbsCat*);
if (abscat2 == NULL) {
gu_raise(err, PgfExn);
GuExnData* exn = gu_raise(err, PgfExn);
exn->data = "Unknown category name";
goto close;
}
if (abscat1->meta_child_probs == NULL) {
abscat1->meta_child_probs =
gu_map_type_new(PgfMetaChildMap, pool);
}
gu_map_put(abscat1->meta_child_probs, abscat2, prob_t, prob);
}
}
close:
gu_pool_free(tmp_pool);
fclose(fp);
}
GuString
pgf_abstract_name(PgfPGF* pgf)
{

5
src/runtime/c/pgf/pgf.h
View File

@@ -80,11 +80,6 @@ pgf_read(const char* fpath,
*
*/
void
pgf_load_meta_child_probs(PgfPGF*, const char* fpath,
GuPool* pool, GuExn* err);
GuString
pgf_abstract_name(PgfPGF*);

2
src/runtime/c/pgf/printer.c
View File

@@ -48,7 +48,7 @@ pgf_print_cat(GuMapItor* fn, const void* key, void* value,
ctxt = next;
}
gu_printf(out, err, " ; -- %f\n",cat->meta_prob);
gu_printf(out, err, " ; -- %f\n", cat->prob);
}
void

8
src/runtime/c/pgf/reader.c
View File

@@ -516,10 +516,6 @@ pgf_read_abscat(PgfReader* rdr, PgfAbstr* abstr, PgfCIdMap* abscats)
gu_return_on_exn(rdr->err, NULL);
}
abscat->meta_prob = INFINITY;
abscat->meta_token_prob = INFINITY;
abscat->meta_child_probs = NULL;
GuBuf* functions = gu_new_buf(PgfAbsFun*, rdr->tmp_pool);
size_t n_functions = pgf_read_len(rdr);
@@ -538,6 +534,8 @@ pgf_read_abscat(PgfReader* rdr, PgfAbstr* abstr, PgfCIdMap* abscats)
gu_buf_push(functions, PgfAbsFun*, absfun);
}
abscat->prob = - log(gu_in_f64be(rdr->in, rdr->err));
pgf_jit_predicate(rdr->jit_state, abscats, abscat, functions);
return abscat;
@@ -1155,6 +1153,8 @@ pgf_read_concrete(PgfReader* rdr, PgfAbstr* abstr, PgfAbsFun* abs_lin_fun)
pgf_read_cid(rdr, rdr->opool);
gu_return_on_exn(rdr->err, NULL);
concr->abstr = abstr;
concr->cflags =
pgf_read_flags(rdr);
gu_return_on_exn(rdr->err, NULL);

19
src/runtime/c/utils/pgf-translate.c
View File

@@ -53,18 +53,17 @@ int main(int argc, char* argv[]) {
// Create the pool that is used to allocate everything
GuPool* pool = gu_new_pool();
int status = EXIT_SUCCESS;
if (argc < 5 || argc > 6) {
fprintf(stderr, "usage: %s pgf cat from-lang to-lang [probs-file]\n", argv[0]);
if (argc < 5) {
fprintf(stderr, "usage: %s pgf cat from-lang to-lang\n", argv[0]);
status = EXIT_FAILURE;
goto fail;
}
char* filename = argv[1];
GuString filename = argv[1];
GuString cat = argv[2];
GuString from_lang = argv[3];
GuString to_lang = argv[4];
// Create an exception frame that catches all errors.
GuExn* err = gu_new_exn(NULL, gu_kind(type), pool);
@@ -78,16 +77,6 @@ int main(int argc, char* argv[]) {
goto fail;
}
if (argc == 6) {
char* meta_probs_filename = argv[5];
pgf_load_meta_child_probs(pgf, meta_probs_filename, pool, err);
if (!gu_ok(err)) {
fprintf(stderr, "Loading meta child probs failed\n");
status = EXIT_FAILURE;
goto fail;
}
}
// Look up the source and destination concrete categories
PgfConcr* from_concr = pgf_get_language(pgf, from_lang);
PgfConcr* to_concr = pgf_get_language(pgf, to_lang);

12
src/runtime/haskell/PGF.hs
View File

@@ -292,8 +292,8 @@ categories pgf = [c | (c,hs) <- Map.toList (cats (abstract pgf))]
categoryContext pgf cat =
case Map.lookup cat (cats (abstract pgf)) of
Just (hypos,_,_) -> Just hypos
Nothing -> Nothing
Just (hypos,_,_,_) -> Just hypos
Nothing -> Nothing
startCat pgf = DTyp [] (lookStartCat pgf) []
@@ -301,8 +301,8 @@ functions pgf = Map.keys (funs (abstract pgf))
functionsByCat pgf cat =
case Map.lookup cat (cats (abstract pgf)) of
Just (_,fns,_) -> map snd fns
Nothing -> []
Just (_,fns,_,_) -> map snd fns
Nothing -> []
functionType pgf fun =
case Map.lookup fun (funs (abstract pgf)) of
@@ -325,8 +325,8 @@ browse pgf id = fmap (\def -> (def,producers,consumers)) definition
in ppCId id <+> hsep ds <+> char '=' <+> ppExpr 0 scope res | Equ patts res <- eqs])
Just (ty,_,Nothing, _,_) -> Just $ render (text "data" <+> ppCId id <+> colon <+> ppType 0 [] ty)
Nothing -> case Map.lookup id (cats (abstract pgf)) of
Just (hyps,_,_) -> Just $ render (text "cat" <+> ppCId id <+> hsep (snd (mapAccumL (ppHypo 4) [] hyps)))
Nothing -> Nothing
Just (hyps,_,_,_) -> Just $ render (text "cat" <+> ppCId id <+> hsep (snd (mapAccumL (ppHypo 4) [] hyps)))
Nothing -> Nothing
(producers,consumers) = Map.foldWithKey accum ([],[]) (funs (abstract pgf))
where

4
src/runtime/haskell/PGF/Binary.hs
View File

@@ -40,13 +40,13 @@ instance Binary CId where
instance Binary Abstr where
put abs = put (aflags abs,
fmap (\(w,x,y,z,_) -> (w,x,y,z)) (funs abs),
fmap (\(x,y,_) -> (x,y)) (cats abs))
fmap (\(x,y,z,_) -> (x,y,z)) (cats abs))
get = do aflags <- get
funs <- get
cats <- get
return (Abstr{ aflags=aflags
, funs=fmap (\(w,x,y,z) -> (w,x,y,z,0)) funs
, cats=fmap (\(x,y) -> (x,y,0)) cats
, cats=fmap (\(x,y,z) -> (x,y,z,0)) cats
, code=BS.empty
})

8
src/runtime/haskell/PGF/Data.hs
View File

@@ -29,10 +29,10 @@ data PGF = PGF {
data Abstr = Abstr {
aflags :: Map.Map CId Literal, -- ^ value of a flag
funs :: Map.Map CId (Type,Int,Maybe [Equation],Double,BCAddr), -- ^ type, arrity and definition of function + probability
cats :: Map.Map CId ([Hypo],[(Double, CId)],BCAddr), -- ^ 1. context of a category
-- ^ 2. functions of a category. The order in the list is important,
-- this is the order in which the type singatures are given in the source.
-- The termination of the exhaustive generation might depend on this.
cats :: Map.Map CId ([Hypo],[(Double, CId)],Double,BCAddr), -- ^ 1. context of a category
-- 2. functions of a category. The functions are stored
-- in decreasing probability order.
-- 3. probability
code :: BS.ByteString
}

4
src/runtime/haskell/PGF/Macros.hs
View File

@@ -67,7 +67,7 @@ functionsToCat :: PGF -> CId -> [(CId,Type)]
functionsToCat pgf cat =
[(f,ty) | (_,f) <- fs, Just (ty,_,_,_,_) <- [Map.lookup f $ funs $ abstract pgf]]
where
(_,fs,_) = lookMap ([],[],0) cat $ cats $ abstract pgf
(_,fs,_,_) = lookMap ([],[],0,0) cat $ cats $ abstract pgf
-- | List of functions that lack linearizations in the given language.
missingLins :: PGF -> Language -> [CId]
@@ -82,7 +82,7 @@ restrictPGF :: (CId -> Bool) -> PGF -> PGF
restrictPGF cond pgf = pgf {
abstract = abstr {
funs = Map.filterWithKey (\c _ -> cond c) (funs abstr),
cats = Map.map (\(hyps,fs,addr) -> (hyps,filter (cond . snd) fs,addr)) (cats abstr)
cats = Map.map (\(hyps,fs,p,addr) -> (hyps,filter (cond . snd) fs,p,addr)) (cats abstr)
}
} ---- restrict concrs also, might be needed
where

4
src/runtime/haskell/PGF/Printer.hs
View File

@@ -26,8 +26,8 @@ ppAbs name a = text "abstract" <+> ppCId name <+> char '{' $$
ppFlag :: CId -> Literal -> Doc
ppFlag flag value = text "flag" <+> ppCId flag <+> char '=' <+> ppLit value <+> char ';'
ppCat :: CId -> ([Hypo],[(Double,CId)],BCAddr) -> Doc
ppCat c (hyps,_,_) = text "cat" <+> ppCId c <+> hsep (snd (mapAccumL (ppHypo 4) [] hyps)) <+> char ';'
ppCat :: CId -> ([Hypo],[(Double,CId)],Double,BCAddr) -> Doc
ppCat c (hyps,_,_,_) = text "cat" <+> ppCId c <+> hsep (snd (mapAccumL (ppHypo 4) [] hyps)) <+> char ';'
ppFun :: CId -> (Type,Int,Maybe [Equation],Double,BCAddr) -> Doc
ppFun f (t,_,Just eqs,_,_) = text "fun" <+> ppCId f <+> colon <+> ppType 0 [] t <+> char ';' $$

25
src/runtime/haskell/PGF/Probabilistic.hs
View File

@@ -24,13 +24,14 @@ import Data.Maybe (fromMaybe) --, fromJust
-- the probabilities for the different functions in a grammar.
data Probabilities = Probs {
funProbs :: Map.Map CId Double,
catProbs :: Map.Map CId [(Double, CId)]
catProbs :: Map.Map CId (Double, [(Double, CId)])
}
-- | Renders the probability structure as string
showProbabilities :: Probabilities -> String
showProbabilities = unlines . map pr . Map.toList . funProbs where
pr (f,d) = showCId f ++ "\t" ++ show d
showProbabilities = unlines . concatMap prProb . Map.toList . catProbs where
prProb (c,(p,fns)) = pr (p,c) : map pr fns
pr (p,f) = showCId f ++ "\t" ++ show p
-- | Reads the probabilities from a file.
-- This should be a text file where on every line
@@ -50,8 +51,12 @@ readProbabilitiesFromFile file pgf = do
-- for the result category.
mkProbabilities :: PGF -> Map.Map CId Double -> Probabilities
mkProbabilities pgf probs =
let funs1 = Map.fromList [(f,p) | (_,cf) <- Map.toList cats1, (p,f) <- cf]
cats1 = Map.map (\(_,fs,_) -> sortBy cmpProb (fill fs)) (cats (abstract pgf))
let funs1 = Map.fromList [(f,p) | (_,(_,fns)) <- Map.toList cats1, (p,f) <- fns]
cats1 = Map.mapWithKey (\c (_,fns,_,_) ->
let p' = fromMaybe 0 (Map.lookup c probs)
fns' = sortBy cmpProb (fill fns)
in (p', fns'))
(cats (abstract pgf))
in Probs funs1 cats1
where
cmpProb (p1,_) (p2,_) = compare p2 p1
@@ -71,15 +76,15 @@ defaultProbabilities pgf = mkProbabilities pgf Map.empty
getProbabilities :: PGF -> Probabilities
getProbabilities pgf = Probs {
funProbs = Map.map (\(_,_,_,p,_) -> p) (funs (abstract pgf)),
catProbs = Map.map (\(_,fns,_) -> fns) (cats (abstract pgf))
funProbs = Map.map (\(_,_,_,p,_) -> p ) (funs (abstract pgf)),
catProbs = Map.map (\(_,fns,p,_) -> (p,fns)) (cats (abstract pgf))
}
setProbabilities :: Probabilities -> PGF -> PGF
setProbabilities probs pgf = pgf {
abstract = (abstract pgf) {
funs = mapUnionWith (\(ty,a,df,_,addr) p -> (ty,a,df,p,addr)) (funs (abstract pgf)) (funProbs probs),
cats = mapUnionWith (\(hypos,_,addr) fns -> (hypos,fns,addr)) (cats (abstract pgf)) (catProbs probs)
funs = mapUnionWith (\(ty,a,df,_,addr) p -> (ty,a,df, p,addr)) (funs (abstract pgf)) (funProbs probs),
cats = mapUnionWith (\(hypos,_,_,addr) (p,fns) -> (hypos,fns,p,addr)) (cats (abstract pgf)) (catProbs probs)
}}
where
mapUnionWith f map1 map2 =
@@ -102,7 +107,7 @@ rankTreesByProbs pgf ts = sortBy (\ (_,p) (_,q) -> compare q p)
mkProbDefs :: PGF -> ([[CId]],[(CId,Type,[Equation])])
mkProbDefs pgf =
let cs = [(c,hyps,fns) | (c,(hyps0,fs,_)) <- Map.toList (cats (abstract pgf)),
let cs = [(c,hyps,fns) | (c,(hyps0,fs,_,_)) <- Map.toList (cats (abstract pgf)),
not (elem c [cidString,cidInt,cidFloat]),
let hyps = zipWith (\(bt,_,ty) n -> (bt,mkCId ('v':show n),ty))
hyps0

8
src/runtime/haskell/PGF/TypeCheck.hs
View File

@@ -121,8 +121,8 @@ runTcM abstr f ms s = unTcM f abstr (\x ms s cp b -> let (es,xs) = cp b
lookupCatHyps :: CId -> TcM s [Hypo]
lookupCatHyps cat = TcM (\abstr k h ms -> case Map.lookup cat (cats abstr) of
Just (hyps,_,_) -> k hyps ms
Nothing -> h (UnknownCat cat))
Just (hyps,_,_,_) -> k hyps ms
Nothing -> h (UnknownCat cat))
lookupFunType :: CId -> TcM s Type
lookupFunType fun = TcM (\abstr k h ms -> case Map.lookup fun (funs abstr) of
@@ -143,8 +143,8 @@ typeGenerators scope cat = fmap normalize (liftM2 (++) x y)
| cat == cidString = return [(1.0,ELit (LStr "Foo"),TTyp [] (DTyp [] cat []))]
| otherwise = TcM (\abstr k h ms ->
case Map.lookup cat (cats abstr) of
Just (_,fns,_) -> unTcM (mapM helper fns) abstr k h ms
Nothing -> h (UnknownCat cat))
Just (_,fns,_,_) -> unTcM (mapM helper fns) abstr k h ms
Nothing -> h (UnknownCat cat))
helper (p,fn) = do
ty <- lookupFunType fn

129702
treebanks/PennTreebank/ParseEngAbs.probs
View File

File diff suppressed because it is too large Load Diff

31
treebanks/PennTreebank/ParseEngAbs3.probs
View File

@@ -1,31 +0,0 @@
Phr * 1e-5
Phr A 4.276438451840496e-3
Phr AP 2.2906320842384386e-3
Phr AdA 2.9594729770522466e-6
Phr AdV 3.5217728426921733e-3
Phr Adv 4.47205961562365e-2
Phr CN 5.918945954104493e-6
Phr Cl 0.23274183333431983
Phr ClSlash 2.0716310839365724e-5
Phr Conj 1.2255177597973352e-2
Phr Det 1.8052785160018703e-4
Phr IAdv 2.9594729770522466e-6
Phr IP 2.9594729770522466e-6
Phr N 8.452254822461217e-3
Phr NP 7.473557108950038e-2
Phr Num 1.1837891908208986e-5
Phr Ord 1.7460890564608255e-4
Phr Predet 2.6635256793470218e-5
Phr Prep 7.56145345636849e-3
Phr Pron 2.3675783816417973e-5
Phr QS 4.143262167873145e-5
Phr Quant 4.03080219474516e-3
Phr RP 6.392461630432853e-4
Phr RS 1.021018177083025e-3
Phr S 5.920425690593019e-2
Phr Subj 6.392461630432853e-4
Phr V 1.071033270395208e-2
Phr V2 1.2784923260865705e-3
Phr VPS 4.751729811955087e-2
Phr VS 1.4797364885261233e-5
Phr _ 1.1965149246222233e-9

93
treebanks/PennTreebank/training.hs
View File

@@ -3,32 +3,31 @@ import qualified Data.Map as Map
import Data.Maybe
import Data.List
grammar_name = "ParseEngAbs.pgf"
treebank_name = "log4.txt"
chunk_cats = map mkCId
["A", "AP", "AdA", "AdV", "Adv", "CN", "Cl", "ClSlash", "Conj", "Det",
"IAdv", "IP", "N", "NP", "Num", "Ord", "Predet", "Prep", "Pron", "QS",
"Quant", "RP", "RS", "S", "Subj", "V", "V2", "VPS", "VS"]
main = do
pgf <- readPGF "ParseEngAbs.pgf"
ls <- fmap (filterExprs . zip [1..] . lines) $ readFile "log4.txt"
pgf <- readPGF grammar_name
ls <- fmap (filterExprs . zip [1..] . lines) $ readFile treebank_name
putStrLn ""
putStrLn ("trees: "++show (length ls))
let stats = foldl' (collectStats pgf)
let (_,cat,_) = unType (startCat pgf)
stats = foldl' (collectStats pgf)
(initStats pgf)
[(n,fromMaybe (error l) (readExpr (toQ l)),Just (mkCId "Phr"),Nothing) | (n,l) <- ls]
[(n,fromMaybe (error l) (readExpr (toQ l)),Just cat) | (n,l) <- ls]
putStrLn ("coverage: "++show (coverage stats))
putStrLn ("Writing ParseEngAbs.probs...")
writeFile "ParseEngAbs.probs" (unlines [show f ++ "\t" ++ show p | (f,p) <- uprobs pgf stats])
putStrLn ("Writing ParseEngAbs2.probs...")
writeFile "ParseEngAbs2.probs" (unlines [show cat1 ++ "\t" ++ show cat2 ++ "\t" ++ show p | (cat1,cat2,p) <- mprobs pgf stats])
putStrLn ("Writing global.probs...")
writeFile "global.probs" (unlines [show f ++ "\t" ++ show p | (f,p) <- gprobs pgf stats])
putStrLn ("Writing categories.probs...")
writeFile "categories.probs" (unlines [show f ++ "\t" ++ show p | (f,p) <- cprobs pgf stats])
writeFile "ParseEngAbs.probs" (unlines [show f ++ "\t" ++ show p | (f,p) <- probs pgf stats])
where
toQ [] = []
toQ ('[':cs) = let (xs,']':ys) = break (==']') cs
in toQ ('?' : ys)
toQ ('?':cs) = 'Q' : toQ cs
toQ (c:cs) = c : toQ cs
@@ -42,31 +41,26 @@ initStats pgf =
(Map.fromListWith (+)
([(f,1) | f <- functions pgf] ++
[(cat pgf f,1) | f <- functions pgf])
,Map.empty
,0
)
collectStats pgf (ustats,bstats,count) (n,e,mb_cat1,mb_cat2) =
collectStats pgf (ustats,count) (n,e,mb_cat1) =
case unApp e of
Just (f,args) -> let fcat2 = cat2 pgf f n e
fcat = fromMaybe (cat2 pgf f n e) mb_cat1
fcat = fromMaybe fcat2 mb_cat1
cf = fromMaybe 0 (Map.lookup f ustats)
cc = fromMaybe 0 (Map.lookup fcat ustats)
in if isJust mb_cat1 && f /= mkCId "Q" && fcat /= fcat2
then error (show n ++ ": " ++ showExpr [] e)
else
cf `seq` cc `seq` bstats `seq` count `seq`
cf `seq` cc `seq` count `seq`
foldl' (collectStats pgf)
(Map.insert f (cf+1) (Map.insert fcat (cc+1) ustats)
,(if null args
then Map.insertWith (+) (fcat,wildCId) 1
else id)
(maybe bstats (\cat2 -> Map.insertWith (+) (cat2,fcat) 1 bstats) mb_cat2)
,count+1
)
(zipWith3 (\e mb_cat1 mb_cat2 -> (n,e,mb_cat1,mb_cat2)) args (argCats f) (repeat (Just fcat)))
(zipWith (\e mb_cat1 -> (n,e,mb_cat1)) args (argCats f))
Nothing -> case unStr e of
Just _ -> (ustats,bstats,count+1)
Just _ -> (ustats,count+1)
Nothing -> error ("collectStats ("++showExpr [] e++")")
where
argCats f =
@@ -75,44 +69,25 @@ collectStats pgf (ustats,bstats,count) (n,e,mb_cat1,mb_cat2) =
in map tyCat arg_tys
Nothing -> repeat Nothing
coverage (ustats,bstats,count) =
let c = fromMaybe 0 (Map.lookup (mkCId "Q") ustats)
in (fromIntegral (count - c) / fromIntegral count) * 100
coverage (stats,gcount) =
let c = fromMaybe 0 (Map.lookup (mkCId "Q") stats)
in (fromIntegral (gcount - c) / fromIntegral gcount) * 100
uprobs pgf (ustats,bstats,count) =
[toProb f (cat pgf f) | f <- functions pgf]
probs pgf (stats,gcount) =
[toFProb f (cat pgf f) | f <- functions pgf] ++
[toCProb c | c <- chunk_cats]
where
toProb f cat =
let count = fromMaybe 0 (Map.lookup f ustats)
cat_mass = fromMaybe 0 (Map.lookup cat ustats)
toFProb f cat =
let count = fromMaybe 0 (Map.lookup f stats)
cat_mass = fromMaybe 0 (Map.lookup cat stats)
in (f, fromIntegral count / fromIntegral cat_mass :: Double)
mprobs pgf (ustats,bstats,count) =
concat [toProb cat | cat <- categories pgf]
where
toProb cat =
let mass = sum [count | ((cat1,cat2),count) <- Map.toList bstats, cat1==cat]
cat_count = fromMaybe 0 (Map.lookup cat ustats)
fun_count = sum [fromMaybe 0 (Map.lookup f ustats) | f <- functionsByCat pgf cat]
in (cat,mkCId "*",if cat_count == 0 then 0 else fromIntegral (cat_count - fun_count) / fromIntegral cat_count) :
[(cat1,cat2,fromIntegral count / fromIntegral mass)
| ((cat1,cat2),count) <- Map.toList bstats, cat1==cat]
toCProb c =
let ccount = fromMaybe 0 (Map.lookup c stats)
in (c, fromIntegral ccount / fromIntegral chunk_mass :: Double)
gprobs pgf (ustats,bstats,count) =
sortBy (\x y -> compare (snd y) (snd x)) [toProb f | f <- functions pgf]
where
toProb f =
let fcount = fromMaybe 0 (Map.lookup f ustats)
in (f, fromIntegral fcount / fromIntegral count :: Double)
cprobs pgf (ustats,bstats,count) =
sortBy (\x y -> compare (snd y) (snd x)) [toProb c | c <- categories pgf]
where
mass = sum [fromMaybe 0 (Map.lookup c ustats) | c <- categories pgf]
toProb c =
let fcount = fromMaybe 0 (Map.lookup c ustats)
in (c, fromIntegral fcount / fromIntegral mass :: Double)
chunk_mass =
sum [fromMaybe 0 (Map.lookup c stats) | c <- chunk_cats]
cat pgf f =
case fmap unType (functionType pgf f) of