% GenI surface realiser % Copyright (C) 2005 Carlos Areces and Eric Kow % % This program is free software; you can redistribute it and/or % modify it under the terms of the GNU General Public License % as published by the Free Software Foundation; either version 2 % of the License, or (at your option) any later version. % % This program is distributed in the hope that it will be useful, % but WITHOUT ANY WARRANTY; without even the implied warranty of % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the % GNU General Public License for more details. % % You should have received a copy of the GNU General Public License % along with this program; if not, write to the Free Software % Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. \chapter{Btypes} \label{cha:Btypes} This module provides basic datatypes like GNode, as well as operations on trees, nodes and semantics. Things here are meant to be relatively low-level and primitive (well, with the exception of feature structure unification, that is). \begin{code}
{-# GHC_OPTIONS -fglasgow-exts #-}
-- {-# LANGUAGE TypeSynonymInstances, FlexibleInstances, DeriveDataTypeable #-}
module NLP.GenI.Btypes(
   -- Datatypes
   GNode(..), GType(Subs, Foot, Lex, Other), NodeName,
   Ttree(..), MTtree, SemPols, TestCase(..),
   Pred, Flist, AvPair, GeniVal(..),
   Lexicon, ILexEntry(..), MorphLexEntry, Macros, Sem, LitConstr, SemInput, Subst,
   emptyLE, emptyGNode, emptyMacro,

   -- GNode stuff
   gCategory, showLexeme, lexemeAttributes, gnnameIs,

   -- Functions from Tree GNode
   plugTree, spliceTree,
   root, rootUpd, foot, setLexeme, setAnchor,

   -- Functions from Sem
   toKeys, subsumeSem, sortSem, showSem, showPred,

   -- Functions from Flist
   sortFlist, unify, unifyFeat, mergeSubst,
   showFlist, showPairs, showAv,

   -- Other functions
   Replacable(..), replaceOneAsMap,
   Collectable(..), Idable(..),
   alphaConvert, alphaConvertById,
   fromGConst, fromGVar,
   isConst, isVar, isAnon,

   -- Polarities

   -- Tests
   prop_unify_anon, prop_unify_self, prop_unify_sym
) where
\end{code} \ignore{ \begin{code}
-- import Debug.Trace -- for test stuff
import Control.Monad (liftM)
import Data.List
import Data.Maybe (fromMaybe, isJust, mapMaybe)
import Data.Generics (Data)
import Data.Typeable (Typeable)
import qualified Data.Map as Map
import qualified Data.Set as Set
import Data.Tree
import Test.QuickCheck hiding (collect) -- needed for testing via ghci

import NLP.GenI.General(map', filterTree, listRepNode, snd3, geniBug, comparing)
--instance Show (IO()) where
--  show _ = ""
\end{code} } % ---------------------------------------------------------------------- \section{Grammar} % ---------------------------------------------------------------------- A grammar is composed of some unanchored trees (macros) and individual lexical entries. The trees are grouped into families. Every lexical entry is associated with a single family. See section section \ref{sec:combine_macros} for the process that combines lexical items and trees into a set of anchored trees. \begin{code}
type MTtree = Ttree GNode
type Macros = [MTtree]

data Ttree a = TT
  { params  :: [GeniVal]
  , pfamily :: String
  , pidname :: String
  , pinterface :: Flist
  , ptype :: Ptype
  , psemantics :: Maybe Sem
  , ptrace :: [String]
  , tree :: Tree a }
  deriving (Show, Data, Typeable)

data Ptype = Initial | Auxiliar | Unspecified
             deriving (Show, Eq, Data, Typeable)

instance (Replacable a) => Replacable (Ttree a) where
  replaceMap s mt =
    mt { params = replaceMap s (params mt)
       , tree   = replaceMap s (tree mt)
       , pinterface  = replaceMap s (pinterface mt)
       , psemantics = replaceMap s (psemantics mt) }
  replaceOne = replaceOneAsMap

instance (Collectable a) => Collectable (Ttree a) where
  collect mt = (collect $ params mt) . (collect $ tree mt) .
               (collect $ psemantics mt) . (collect $ pinterface mt)

-- | A null tree which you can use for various debugging or display purposes.
emptyMacro :: MTtree
emptyMacro = TT { params  = [],
                  pidname = "",
                  pfamily = "",
                  pinterface = [],
                  ptype = Unspecified,
                  psemantics = Nothing,
                  ptrace = [],
                  tree  = Node emptyGNode []
\end{code} \paragraph{Lexical entries} \begin{code}
-- | A lexicon maps semantic predicates to lexical entries.
type Lexicon = Map.Map String [ILexEntry]
type SemPols  = [Int]
data ILexEntry = ILE
    { -- normally just a singleton, useful for merging synonyms
      iword       :: [String]
    , ifamname    :: String
    , iparams     :: [GeniVal]
    , iinterface  :: Flist
    , ifilters    :: Flist
    , iequations  :: Flist
    , iptype      :: Ptype
    , isemantics  :: Sem
    , isempols    :: [SemPols] }
  deriving (Show, Eq, Data, Typeable)

instance Replacable ILexEntry where
  replaceMap s i =
    i { iinterface  = replaceMap s (iinterface i)
      , iequations  = replaceMap s (iequations i)
      , isemantics  = replaceMap s (isemantics i)
      , iparams = replaceMap s (iparams i) }
  replaceOne = replaceOneAsMap

instance Collectable ILexEntry where
  collect l = (collect $ iinterface l) . (collect $ iparams l) .
              (collect $ ifilters l) . (collect $ iequations l) .
              (collect $ isemantics l)

emptyLE :: ILexEntry
emptyLE = ILE { iword = [],
                ifamname = "",
                iparams = [],
                iinterface   = [],
                ifilters = [],
                iptype = Unspecified,
                isemantics = [],
                iequations = [],
                isempols   = [] }
\end{code} \begin{code}
type MorphLexEntry = (String,String,Flist)
\end{code} % ---------------------------------------------------------------------- \section{GNode} % ---------------------------------------------------------------------- A GNode is a single node of a syntactic tree. It has a name (gnname), top and bottom feature structures (gup, gdown), a lexeme (ganchor, glexeme: False and empty string if n/a), and some flags information (gtype, gaconstr). \begin{code}
data GNode = GN{gnname :: NodeName,
                gup    :: Flist,
                gdown  :: Flist,
                ganchor  :: Bool,
                glexeme  :: [String],
                gtype    :: GType,
                gaconstr :: Bool,
                gorigin  :: String  -- ^ for TAG, this would be the elementary tree
                                    --   that this node originally came from
           deriving (Eq, Data, Typeable)

-- Node type used during parsing of the grammar
data GType = Subs | Foot | Lex | Other
           deriving (Show, Eq, Data, Typeable)

type NodeName = String

-- | A null 'GNode' which you can use for various debugging or display purposes.
emptyGNode :: GNode
emptyGNode = GN { gnname = "",
                  gup = [], gdown = [],
                  ganchor = False,
                  glexeme = [],
                  gtype = Other,
                  gaconstr = False,
                  gorigin = "" }

gnnameIs :: NodeName -> GNode -> Bool
gnnameIs n = (== n) . gnname
\end{code} A TAG node may have a category. In the core GenI algorithm, there is nothing which distinguishes the category from any other attributes. But for some other uses, such as checking if it is a result or for display purposes, we do treat this attribute differently. We take here the convention that the category of a node is associated to the attribute ``cat''. \begin{code}
-- | Return the value of the "cat" attribute, if available
gCategory :: Flist -> Maybe GeniVal
gCategory top =
  case [ v | (a,v) <- top, a == "cat" ] of
  []  -> Nothing
  [c] -> Just c
  _   -> geniBug $ "Impossible case: node with more than one category"
\end{code} A TAG node might also have a lexeme. If we are lucky, this is explicitly set in the glexeme field of the node. Otherwise, we try to guess it from a list of distinguished attributes (in order of preference). \begin{code}
-- | Attributes recognised as lexemes, in order of preference
lexemeAttributes :: [String]
lexemeAttributes = [ "lex", "phon", "cat" ]
\end{code} \paragraph{show (GNode)} the default show for GNode tries to be very compact; it only shows the value for cat attribute and any flags which are marked on that node. \begin{code}
instance Show GNode where
  show gn =
    let cat_ = case gCategory.gup $ gn of
               Nothing -> []
               Just c  -> show c
        lex_ = showLexeme $ glexeme gn
        stub = concat $ intersperse ":" $ filter (not.null) [ cat_, lex_ ]
        extra = case (gtype gn) of
                   Subs -> " !"
                   Foot -> " *"
                   _    -> if (gaconstr gn)  then " #"   else ""
    in stub ++ extra

-- FIXME: will have to think of nicer way - one which involves
-- unpacking the trees :-(
showLexeme :: [String] -> String
showLexeme []   = ""
showLexeme [l]  = l
showLexeme xs   = concat $ intersperse "|" xs
\end{code} A Replacement on a GNode consists of replacements on its top and bottom feature structures \begin{code}
instance Replacable GNode where
  replaceOne s gn =
    gn { gup = replaceOne s (gup gn)
       , gdown = replaceOne s (gdown gn) }
  replaceMap s gn =
    gn { gup = replaceMap s (gup gn)
       , gdown = replaceMap s (gdown gn) }
\end{code} % ---------------------------------------------------------------------- \section{Tree manipulation} % ---------------------------------------------------------------------- \begin{code}
instance (Replacable a) => Replacable (Tree a) where
  replaceOne s t = fmap (replaceOne s) t
  replaceMap s t = fmap (replaceMap s) t
\end{code} Projector and Update function for Tree \begin{code}
root :: Tree a -> a
root (Node a _) = a

rootUpd :: Tree a -> a -> Tree a
rootUpd (Node _ l) b = (Node b l)

foot :: Tree GNode -> GNode
foot t = case filterTree (\n -> gtype n == Foot) t of
         [x] -> x
         _   -> geniBug $ "foot returned weird result"
\end{code} \begin{code}
-- | Given a lexical item @s@ and a Tree GNode t, returns the tree t'
--   where l has been assigned to the anchor node in t'
setAnchor :: [String] -> Tree GNode -> Tree GNode
setAnchor s t =
  let filt (Node a []) = (gtype a == Lex && ganchor a)
      filt _ = False
  in case listRepNode (setLexeme s) filt [t] of
     ([r],True) -> r
     _ -> geniBug $ "setLexeme " ++ show s ++ " returned weird result"

-- | Given a lexical item @l@ and a tree node @n@ (actually a subtree
--   with no children), return the same node with the lexical item as
--   its unique child.  The idea is that it converts terminal lexeme nodes
--   into preterminal nodes where the actual terminal is the given lexical
--   item
setLexeme :: [String] -> Tree GNode -> Tree GNode
setLexeme l (Node a []) = Node a [ Node subanc [] ]
  where subanc = emptyGNode { gnname = '_' : ((gnname a) ++ ('.' : (concat l)))
                            , gaconstr = True
                            , glexeme = l}
setLexeme _ _ = geniBug "impossible case in setLexeme - subtree with kids"
\end{code} \subsection{Substitution and Adjunction} This module handles the strictly syntactic part of the TAG substitution and adjunction. We do substitution with a very general \fnreflite{plugTree} function, whose only job is to plug two trees together at a specified node. Note that this function is also used to implement adjunction. \begin{code}
-- | Plug the first tree into the second tree at the specified node.
--   Anything below the second node is silently discarded.
--   We assume the trees are pluggable; it is treated as a bug if
--   they are not!
plugTree :: Tree NodeName -> NodeName -> Tree NodeName -> Tree NodeName
plugTree male n female =
  case listRepNode (const male) (nmatch n) [female] of
  ([r], True) -> r
  _           -> geniBug $ "unexpected plug failure at node " ++ n

-- | Given two trees 'auxt' and 't', splice the tree 'auxt' into
--   't' via the TAG adjunction rule.
spliceTree :: NodeName      -- ^ foot node of the aux tree
           -> Tree NodeName -- ^ aux tree
           -> NodeName      -- ^ place to adjoin in target tree
           -> Tree NodeName -- ^ target tree
           -> Tree NodeName
spliceTree f auxT n targetT =
  case findSubTree n targetT of -- excise the subtree at n
  Nothing -> geniBug $ "Unexpected adjunction failure. " ++
                       "Could not find node " ++ n ++ " of target tree."
  Just eT -> -- plug the excised bit into the aux
             let auxPlus = plugTree eT f auxT
             -- plug the augmented aux at n
             in  plugTree auxPlus n targetT

nmatch :: NodeName -> Tree NodeName -> Bool
nmatch n (Node a _) = a == n

findSubTree :: NodeName -> Tree NodeName -> Maybe (Tree NodeName)
findSubTree n n2@(Node x ks)
  | x == n    = Just n2
  | otherwise = case mapMaybe (findSubTree n) ks of
                []    -> Nothing
                (h:_) -> Just h
\end{code} % ---------------------------------------------------------------------- \section{Features and variables} % ---------------------------------------------------------------------- \begin{code}
type Flist   = [AvPair]
type AvPair  = (String,GeniVal)
\end{code} \subsection{GeniVal} \begin{code}
data GeniVal = GConst [String]
             | GVar   !String
             | GAnon
  deriving (Eq,Ord, Data, Typeable)

instance Show GeniVal where
  show (GConst x) = concat $ intersperse "|" x
  show (GVar x)   = '?':x
  show GAnon      = "?_"

isConst :: GeniVal -> Bool
isConst (GConst _) = True
isConst _ = False

isVar :: GeniVal -> Bool
isVar (GVar _) = True
isVar _        = False

isAnon :: GeniVal -> Bool
isAnon GAnon = True
isAnon _     = False

-- | (assumes that it's a GConst!)
fromGConst :: GeniVal -> [String]
fromGConst (GConst x) = x
fromGConst x = error ("fromGConst on " ++ show x)

-- | (assumes that it's a GVar!)
fromGVar :: GeniVal -> String
fromGVar (GVar x) = x
fromGVar x = error ("fromGVar on " ++ show x)
\end{code} \subsection{Collectable} A Collectable is something which can return its variables as a set. By variables, what I most had in mind was the GVar values in a GeniVal. This notion is probably not very useful outside the context of alpha-conversion task, but it seems general enough that I'll keep it around for a good bit, until either some use for it creeps up, or I find a more general notion that I can transform this into. \begin{code}
class Collectable a where
  collect :: a -> Set.Set String -> Set.Set String

instance Collectable a => Collectable (Maybe a) where
  collect Nothing  s = s
  collect (Just x) s = collect x s

instance (Collectable a => Collectable [a]) where
  collect l s = foldr collect s l

instance (Collectable a => Collectable (Tree a)) where
  collect = collect.flatten

-- Pred is what I had in mind here
instance ((Collectable a, Collectable b, Collectable c)
           => Collectable (a,b,c)) where
  collect (a,b,c) = collect a . collect b . collect c

instance Collectable GeniVal where
  collect (GVar v) s = Set.insert v s
  collect _ s = s

instance Collectable (String,GeniVal) where
  collect (_,b) = collect b

instance Collectable GNode where
  collect n = (collect $ gdown n) . (collect $ gup n)
\end{code} \subsection{Replacable} \label{sec:replacable} \label{sec:replacements} The idea of replacing one variable value with another is something that appears all over the place in GenI. So we try to smooth out its use by making a type class out of it. \begin{code}
class Replacable a where
  replace :: Subst -> a -> a
  replace = replaceMap

  replaceMap :: Map.Map String GeniVal -> a -> a

  replaceOne :: (String,GeniVal) -> a -> a

  -- | Here it is safe to say (X -> Y; Y -> Z) because this would be crushed
  --   down into a final value of (X -> Z; Y -> Z)
  replaceList :: [(String,GeniVal)] -> a -> a
  replaceList = replaceMap . foldl update Map.empty
     update m (s1,s2) = Map.insert s1 s2 $ Map.map (replaceOne (s1,s2)) m

-- | Default implementation for replaceOne but not a good idea for the
--   core stuff; which is why it is not a typeclass default
replaceOneAsMap :: Replacable a => (String, GeniVal) -> a -> a
replaceOneAsMap s = replaceMap (uncurry Map.singleton s)

instance (Replacable a => Replacable (Maybe a)) where
  replaceMap s = liftM (replaceMap s)
  replaceOne s = liftM (replaceOne s)
\end{code} GeniVal is probably the simplest thing you would one to apply a substitution on \begin{code}
instance Replacable GeniVal where
  replaceMap m v@(GVar v_) = {-# SCC "replaceMap" #-} Map.findWithDefault v v_ m
  replaceMap _ v = {-# SCC "replaceMap" #-} v

  replaceOne (s1, s2) (GVar v_) | v_ == s1 = {-# SCC "replaceOne" #-} s2
  replaceOne _ v = {-# SCC "replaceOne" #-} v
\end{code} Substitution on list consists of performing substitution on each item. Each item, is independent of the other, of course. \begin{code}
instance (Replacable a => Replacable [a]) where
  replaceMap s = {-# SCC "replaceMap" #-} map' (replaceMap s)
  replaceOne s = {-# SCC "replaceOne" #-} map' (replaceOne s)
\end{code} Substitution on an attribute/value pairs consists of ignoring the attribute and performing substitution on the value. \begin{code}
instance Replacable AvPair where
  replaceMap s (a,v) = {-# SCC "replaceMap" #-} (a, replaceMap s v)
  replaceOne s (a,v) = {-# SCC "replaceOne" #-} (a, replaceOne s v)

instance Replacable (String, ([String], Flist)) where
  replaceMap s (n,(a,v)) = {-# SCC "replaceMap" #-} (n,(a, replaceMap s v))
  replaceOne s (n,(a,v)) = {-# SCC "replaceOne" #-} (n,(a, replaceOne s v))
\end{code} \subsection{Idable} An Idable is something that can be mapped to a unique id. You might consider using this to implement Ord, but I won't. Note that the only use I have for this so far (20 dec 2005) is in alpha-conversion. \begin{code}
class Idable a where
  idOf :: a -> Integer
\end{code} \subsection{Other feature and variable stuff} Our approach to $\alpha$-conversion works by appending a unique suffix to all variables in an object. See section \ref{sec:fs_unification} for why we want this. \begin{code}
alphaConvertById :: (Collectable a, Replacable a, Idable a) => a -> a
alphaConvertById x = {-# SCC "alphaConvertById" #-}
  alphaConvert ('-' : (show . idOf $ x)) x

alphaConvert :: (Collectable a, Replacable a) => String -> a -> a
alphaConvert suffix x = {-# SCC "alphaConvert" #-}
  let vars   = Set.elems $ collect x Set.empty
      convert v = GVar (v ++ suffix)
      subst = Map.fromList $ map (\v -> (v, convert v)) vars
  in replaceMap subst x
\end{code} \begin{code}
-- | Sort an Flist according with its attributes
sortFlist :: Flist -> Flist
sortFlist = sortBy (comparing fst)

showFlist :: Flist -> String
showFlist f = "[" ++ showPairs f ++ "]"

showPairs :: Flist -> String
showPairs = unwords . map showAv

showAv :: AvPair -> String
showAv (y,z) = y ++ ":" ++ show z
\end{code} % ---------------------------------------------------------------------- \section{Semantics} \label{btypes_semantics} % ---------------------------------------------------------------------- \begin{code}
-- handle, predicate, parameters
type Pred = (GeniVal, GeniVal, [GeniVal])
type Sem = [Pred]
type LitConstr = (Pred, [String])
type SemInput  = (Sem,Flist,[LitConstr])
type Subst = Map.Map String GeniVal

data TestCase = TestCase
       { tcName :: String
       , tcSemString :: String -- ^ for gui
       , tcSem  :: SemInput
       , tcExpected :: [String] -- ^ expected results (for testing)
       , tcOutputs :: [(String, Map.Map (String,String) [String])]
       -- ^ results we actually got, and their traces (for testing)
       } deriving Show

emptyPred :: Pred
emptyPred = (GAnon,GAnon,[])
\end{code} A replacement on a predicate is just a replacement on its parameters \begin{code}
instance Replacable Pred where
  replaceMap s (h, n, lp) = (replaceMap s h, replaceMap s n, replaceMap s lp)
  replaceOne s (h, n, lp) = (replaceOne s h, replaceOne s n, replaceOne s lp)
\end{code} \begin{code}
showSem :: Sem -> String
showSem l =
    "[" ++ (unwords $ map showPred l) ++ "]"

showPred :: Pred -> String
showPred (h, p, l) = showh ++ show p ++ "(" ++ unwords (map show l) ++ ")"
    hideh (GConst [x]) = "genihandle" `isPrefixOf` x
    hideh _ = False
    showh = if (hideh h) then "" else (show h) ++ ":"
\end{code} \begin{code}
-- | Given a Semantics, return the string with the proper keys
--   (propsymbol+arity) to access the agenda
toKeys :: Sem -> [String]
toKeys l = map (\(_,prop,par) -> show prop ++ (show $ length par)) l
\end{code} \subsection{Semantic subsumption} \label{fn:subsumeSem} FIXME: comment fix Given tsem the input semantics, and lsem the semantics of a potential lexical candidate, returns a list of possible ways that the lexical semantics could subsume the input semantics. We return a pair with the semantics that would result from unification\footnote{We need to do this because there may be anonymous variables}, and the substitutions that need to be propagated throughout the rest of the lexical item later on. Note: we return more than one possible substitution because s could be different subsets of ts. Consider, for example, \semexpr{love(j,m), name(j,john), name(m,mary)} and the candidate \semexpr{name(X,Y)}. TODO WE ASSUME BOTH SEMANTICS ARE ORDERED and that the input semantics is non-empty. \begin{code}
subsumeSem :: Sem -> Sem -> [(Sem,Subst)]
subsumeSem tsem lsem =
  subsumeSemHelper ([],Map.empty) (reverse tsem) (reverse lsem)
\end{code} This is tricky because each substep returns multiple results. We solicit the help of accumulators to keep things from getting confused. \begin{code}
subsumeSemHelper :: (Sem,Subst) -> Sem -> Sem -> [(Sem,Subst)]
subsumeSemHelper _ [] _  =
  error "input semantics is non-empty in subsumeSemHelper"
subsumeSemHelper acc _ []      = [acc]
subsumeSemHelper acc tsem (hd:tl) =
  let (accSem,accSub) = acc
      -- does the literal hd subsume the input semantics?
      pRes = subsumePred tsem hd
      -- toPred reconstructs the literal hd with new parameters p.
      -- The head of the list is taken to be the handle.
      toPred p = (head p, snd3 hd, tail p)
      -- next adds a result from predication subsumption to
      -- the accumulators and goes to the next recursive step
      next (p,s) = subsumeSemHelper acc2 tsem2 tl2
         where tl2   = replace s tl
               tsem2 = replace s tsem
               acc2  = (toPred p : accSem, mergeSubst accSub s)
  in concatMap next pRes
\end{code} \fnlabel{subsumePred} The first Sem s1 and second Sem s2 are the same when we start we circle on s2 looking for a match for Pred, and meanwhile we apply the partical substitutions to s1. Note: we treat the handle as if it were a parameter. \begin{code}
subsumePred :: Sem -> Pred -> [([GeniVal],Subst)]
subsumePred [] _ = []
subsumePred ((h1, p1, la1):l) (pred2@(h2,p2,la2)) =
    -- if we found the proper predicate
    if ((p1 == p2) && (length la1 == length la2))
    then let mrs  = unify (h1:la1) (h2:la2)
             next = subsumePred l pred2
         in maybe next (:next) mrs
    else if (p1 < p2) -- note that the semantics have to be reversed!
         then []
         else subsumePred l pred2
\end{code} \subsection{Other semantic stuff} \begin{code}
-- | Sort semantics first according to its predicate, and then to its handles.
sortSem :: Sem -> Sem
sortSem = sortBy (\(h1,p1,a1) (h2,p2,a2) -> compare (p1, h1:a1) (p2, h2:a2))
\end{code} % -------------------------------------------------------------------- \subsection{Unification} \label{sec:fs_unification} % -------------------------------------------------------------------- Feature structure unification takes two feature lists as input. If it fails, it returns Nothing. Otherwise, it returns a tuple with: \begin{enumerate} \item a unified feature structure list \item a list of variable replacements that will need to be propagated across other feature structures with the same variables \end{enumerate} Unification fails if, at any point during the unification process, the two lists have different constant values for the same attribute. For example, unification fails on the following inputs because they have different values for the \textit{number} attribute: \begin{quotation} \fs{\it cat:np\\ \it number:3\\} \fs{\it cat:np\\ \it number:2\\} \end{quotation} Note that the following input should also fail as a result on the coreference on \textit{?X}. \begin{quotation} \fs{\it cat:np\\ \it one: 1\\ \it two:2\\} \fs{\it cat:np\\ \it one: ?X\\ \it two:?X\\} \end{quotation} On the other hand, any other pair of feature lists should unify succesfully, even those that do not share the same attributes. Below are some examples of successful unifications: \begin{quotation} \fs{\it cat:np\\ \it one: 1\\ \it two:2\\} \fs{\it cat:np\\ \it one: ?X\\ \it two:?Y\\} $\rightarrow$ \fs{\it cat:np\\ \it one: 1\\ \it two:2\\}, \end{quotation} \begin{quotation} \fs{\it cat:np\\ \it number:3\\} \fs{\it cat:np\\ \it case:nom\\} $\rightarrow$ \fs{\it cat:np\\ \it case:nom\\ \it number:3\\}, \end{quotation} \fnlabel{unifyFeat} is an implementation of feature structure unification. It makes the following assumptions: \begin{itemize} \item Features are ordered \item The Flists do not share variables!!! More precisely, if the two Flists have the same variable, they will have the same value. Though this behaviour may not be desirable, we don't really care because we never encounter the situation (see page \pageref{par:lexSelection}). \end{itemize} \begin{code}
unifyFeat :: (Monad m) => Flist -> Flist -> m (Flist, Subst)
unifyFeat f1 f2 =
  {-# SCC "unification" #-}
  let (att, val1, val2) = alignFeat f1 f2
  in att `seq`
     do (res, subst) <- unify val1 val2
        return (zip att res, subst)
\end{code} \fnlabel{alignFeat} The less trivial case is when neither list is empty. If we are looking at the same attribute, then we transfer control to the helper function. Otherwise, we remove the (alphabetically) smaller att-val pair, add it to the results, and move on. This only works if the lists are alphabetically sorted beforehand! \begin{code}
alignFeat :: Flist -> Flist -> ([String], [GeniVal], [GeniVal])
alignFeat [] [] = ([], [], [])

alignFeat [] ((f,v):x) =
  case alignFeat [] x of
  (att, left, right) -> (f:att, GAnon:left, v:right)

alignFeat x [] =
  case alignFeat [] x of
  (att, left, right) -> (att, right, left)

alignFeat fs1@((f1, v1):l1) fs2@((f2, v2):l2)
   | f1 == f2  = case alignFeat l1 l2 of
                 (att, left, right) -> (f1:att, v1:left,    v2:right)
   | f1 <  f2  = case alignFeat l1 fs2 of
                 (att, left, right) -> (f1:att, v1:left, GAnon:right)
   | f1 >  f2  = case alignFeat fs1 l2 of
                 (att, left, right) -> (f2:att, GAnon:left, v2:right)
   | otherwise = error "Feature structure unification is badly broken"
\end{code} \subsection{Unification} \fnlabel{unify} performs unification on two lists of GeniVal. If unification succeeds, it returns \verb!Just (r,s)! where \verb!r! is the result of unification and \verb!s! is a list of substitutions that this unification results in. Notes: \begin{itemize} \item there may be multiple results because of disjunction \item we need to return \verb!r! because of anonymous variables \item the lists need not be same length; we just assume you want the longer of the two \end{itemize} The core unification algorithm follows these rules in order: \begin{enumerate} \item if either h1 or h2 are anonymous, we add the other to the result, and we don't add any replacements. \item if h1 is a variable then we replace it by h2, regardless of whether or not h2 is a variable \item if h2 is a variable then we replace it by h1 \item if neither h1 and h2 are variables, but they match, we arbitarily add one of them to the result, but we don't add any replacements. \item if neither are variables and they do \emph{not} match, we fail \end{enumerate} \begin{code}
unify :: (Monad m) => [GeniVal] -> [GeniVal] -> m ([GeniVal], Subst)
unify [] l2 = {-# SCC "unification" #-} return (l2, Map.empty)
unify l1 [] = {-# SCC "unification" #-} return (l1, Map.empty)
unify (h1:t1) (h2:t2) | h1 == h2 = {-# SCC "unification" #-} unifySansRep h1 t1 t2
unify (GAnon:t1) (h2:t2) = {-# SCC "unification" #-} unifySansRep h2 t1 t2
unify (h1:t1) (GAnon:t2) = {-# SCC "unification" #-} unifySansRep h1 t1 t2
unify (h1@(GVar _):t1) (h2:t2) = {-# SCC "unification" #-} unifyWithRep h1 h2 t1 t2
unify (h1:t1) (h2@(GVar _):t2) = {-# SCC "unification" #-} unifyWithRep h2 h1 t1 t2
-- special cases for efficiency only
unify ((GConst [_]):_) ((GConst [_]):_) = {-# SCC "unification" #-}
  fail "unification failure"
-- end special efficiency-only cases
unify ((GConst h1v):t1) ((GConst h2v):t2) = {-# SCC "unification" #-}
  case h1v `intersect` h2v of
  []   -> fail "unification failure"
  newH -> unifySansRep (GConst newH) t1 t2
{-# INLINE unifySansRep #-}
{-# INLINE unifyWithRep #-}
unifySansRep :: (Monad m) => GeniVal -> [GeniVal] -> [GeniVal] -> m ([GeniVal], Subst)
unifySansRep x2 t1 t2 = {-# SCC "unification" #-}
 do (res,subst) <- unify t1 t2
    return (x2:res, subst)

unifyWithRep :: (Monad m) => GeniVal -> GeniVal -> [GeniVal] -> [GeniVal] -> m ([GeniVal], Subst)
unifyWithRep (GVar h1) x2 t1 t2 = {-# SCC "unification" #-}
 let s = (h1,x2)
     t1_ = replaceOne s t1
     t2_ = replaceOne s t2
     ustep = unify t1_ t2_
 in s `seq` t1_ `seq` t2_ `seq` ustep `seq`
    (ustep >>= \(res,subst) -> return (x2:res, prependToSubst s subst))
unifyWithRep _ _ _ _ = geniBug "unification error"
\end{code} \begin{code}
-- | Note that the first Subst is assumed to come chronologically
--   before the second one; so merging @{ X -> Y }@ and @{ Y -> 3 }@
--   should give us @{ X -> 3; Y -> 3 }@;
--   See 'prependToSubst' for a warning!
mergeSubst :: Subst -> Subst -> Subst
mergeSubst sm1 sm2 = Map.foldWithKey (curry prependToSubst) sm2 sm1

-- | Add to variable replacement to a 'Subst' that logical comes before
--   the other stuff in it.  So for example, if we have @Y -> foo@
--   and we want to insert @X -> Y@, we notice that, in fact, @Y@ has
--   already been replaced by @foo@, so we add @X -> foo@ instead
--   Note that it is undefined if you try to append something like
--   @Y -> foo@ to @Y -> bar@, because that would mean that unification
--   is broken
prependToSubst :: (String,GeniVal) -> Subst -> Subst
prependToSubst (v, gr@(GVar r)) sm
  | isJust $ Map.lookup v sm = geniBug $ "prependToSubst: Eric broke unification.  Prepending " ++ v ++ " twice."
  | otherwise = Map.insert v gr2 sm
  where gr2 = fromMaybe gr $ Map.lookup r sm
prependToSubst (v, gr) sm = Map.insert v gr sm
\end{code} \subsubsection{Unification tests} The unification algorithm should satisfy the following properties: Unifying something with itself should always succeed \begin{code}
prop_unify_self :: [GeniVal] -> Property
prop_unify_self x =
  (all qc_not_empty_GConst) x ==>
    case unify x x of
    Nothing  -> False
    Just unf -> (fst unf == x)
\end{code} Unifying something with only anonymous variables should succeed. \begin{code}
prop_unify_anon :: [GeniVal] -> Bool
prop_unify_anon x =
  case (unify x y) of
    Nothing  -> False
    Just unf -> (fst unf == x)
  where --
    y  = take (length x) $ repeat GAnon
\end{code} Unification should be symmetrical. We can't guarantee these if there are cases where there are variables in the same place on both sides, so we normalise the sides so that this doesn't happen. \begin{code}
prop_unify_sym :: [GeniVal] -> [GeniVal] -> Property
prop_unify_sym x y =
  let u1 = (unify x y) :: Maybe ([GeniVal],Subst)
      u2 = unify y x
      notOverlap (GVar _, GVar _) = False
      notOverlap _ = True
  in (all qc_not_empty_GConst) x &&
     (all qc_not_empty_GConst) y &&
     all (notOverlap) (zip x y) ==> u1 == u2
\end{code} \ignore{ \begin{code}
-- Definition of Arbitrary GeniVal for QuickCheck
newtype GTestString = GTestString String
newtype GTestString2 = GTestString2 String

fromGTestString :: GTestString -> String
fromGTestString (GTestString s) = s

fromGTestString2 :: GTestString2 -> String
fromGTestString2 (GTestString2 s) = s

instance Arbitrary GTestString where
  arbitrary =
    oneof $ map (return . GTestString) $
    [ "a", "apple" , "b", "banana", "c", "carrot", "d", "durian"
    , "e", "eggplant", "f", "fennel" , "g", "grape" ]
  coarbitrary = error "no implementation of coarbitrary for GTestString"

instance Arbitrary GTestString2 where
  arbitrary =
    oneof $ map (return . GTestString2) $
    [ "X", "Y", "Z", "H", "I", "J", "P", "Q", "R", "S", "T", "U"  ]
  coarbitrary = error "no implementation of coarbitrary for GTestString2"

instance Arbitrary GeniVal where
  arbitrary = oneof [ return $ GAnon,
                      liftM (GVar . fromGTestString2) arbitrary,
                      liftM (GConst . nub . sort . map fromGTestString) arbitrary ]
  coarbitrary = error "no implementation of coarbitrary for GeniVal"

qc_not_empty_GConst :: GeniVal -> Bool
qc_not_empty_GConst (GConst []) = False
qc_not_empty_GConst _ = True
\end{code} }