an API for oracles in the GF parser

This commit is contained in:
krasimir
2016-05-10 17:11:39 +00:00
parent c3cb14e741
commit 7cb772ffdb
5 changed files with 266 additions and 36 deletions

View File

@@ -17,7 +17,8 @@ module PGF2 (-- * CId
-- * PGF
PGF,readPGF,AbsName,abstractName,Cat,startCat,categories,
-- * Concrete syntax
ConcName,Concr,languages,parse,parseWithHeuristics,
ConcName,Concr,languages,parse,
parseWithHeuristics, parseWithOracle,
hasLinearization,linearize,linearizeAll,alignWords,
-- * Types
Type(..), Hypo, BindType(..), showType, functionType,
@@ -340,6 +341,88 @@ mkCallbacksMap concr callbacks pool = do
predict_callback _ _ _ _ = return nullPtr
-- | The oracle is a triple of functions.
-- The first two take a category name and a linearization field name
-- and they should return True/False when the corresponding
-- prediction or completion is appropriate. The third function
-- is the oracle for literals.
type Oracle = (Maybe (Cat -> String -> Int -> Bool)
,Maybe (Cat -> String -> Int -> Bool)
,Maybe (Cat -> String -> Int -> Maybe (Expr,Float,Int))
)
parseWithOracle :: Concr -- ^ the language with which we parse
-> Cat -- ^ the start category
-> String -- ^ the input sentence
-> Oracle
-> Either String [(Expr,Float)]
parseWithOracle lang cat sent (predict,complete,literal) =
unsafePerformIO $
do parsePl <- gu_new_pool
exprPl <- gu_new_pool
exn <- gu_new_exn parsePl
enum <- withCString cat $ \cat ->
withCString sent $ \sent -> do
predictPtr <- maybe (return nullFunPtr) (wrapOracleCallback . oracleWrapper) predict
completePtr <- maybe (return nullFunPtr) (wrapOracleCallback . oracleWrapper) complete
literalPtr <- maybe (return nullFunPtr) (wrapOracleLiteralCallback . oracleLiteralWrapper) literal
cback <- hspgf_new_oracle_callback predictPtr completePtr literalPtr parsePl
pgf_parse_with_oracle (concr lang) cat sent cback exn parsePl exprPl
failed <- gu_exn_is_raised exn
if failed
then do is_parse_error <- gu_exn_caught exn gu_exn_type_PgfParseError
if is_parse_error
then do c_tok <- (#peek GuExn, data.data) exn
tok <- peekCString c_tok
gu_pool_free parsePl
gu_pool_free exprPl
return (Left tok)
else do is_exn <- gu_exn_caught exn gu_exn_type_PgfExn
if is_exn
then do c_msg <- (#peek GuExn, data.data) exn
msg <- peekCString c_msg
gu_pool_free parsePl
gu_pool_free exprPl
throwIO (PGFError msg)
else do gu_pool_free parsePl
gu_pool_free exprPl
throwIO (PGFError "Parsing failed")
else do parseFPl <- newForeignPtr gu_pool_finalizer parsePl
exprFPl <- newForeignPtr gu_pool_finalizer exprPl
exprs <- fromPgfExprEnum enum parseFPl (lang,exprFPl)
return (Right exprs)
where
oracleWrapper oracle _ catPtr lblPtr offset = do
cat <- peekCString catPtr
lbl <- peekCString lblPtr
return (oracle cat lbl (fromIntegral offset))
oracleLiteralWrapper oracle _ catPtr lblPtr poffset out_pool = do
cat <- peekCString catPtr
lbl <- peekCString lblPtr
offset <- peek poffset
case oracle cat lbl (fromIntegral offset) of
Just (e,prob,offset) ->
do poke poffset (fromIntegral offset)
-- here we copy the expression to out_pool
c_e <- withGuPool $ \tmpPl -> do
exn <- gu_new_exn tmpPl
(sb,out) <- newOut tmpPl
let printCtxt = nullPtr
pgf_print_expr (expr e) printCtxt 1 out exn
c_str <- gu_string_buf_freeze sb tmpPl
guin <- gu_string_in c_str tmpPl
pgf_read_expr guin out_pool exn
ep <- gu_malloc out_pool (#size PgfExprProb)
(#poke PgfExprProb, expr) ep c_e
(#poke PgfExprProb, prob) ep prob
return ep
Nothing -> do return nullPtr
hasLinearization :: Concr -> Fun -> Bool
hasLinearization lang id = unsafePerformIO $
withCString id (pgf_has_linearization (concr lang))