• introduction to GF
• introduction to the Resource Grammar Library
• tools
• applications
• ???

• Kaarel Kaljurand
  • https://plus.google.com/111912457138686732050/about
  • https://github.com/Kaljurand
• post doc @ Uni Zurich
  • working in MOLTO on multilingual CNL-based semantic wiki
• GF experience: 1.5 years
• GF projects
  • multilingual ACE in GF
  • Estonian speech recognition grammars
  • very preliminary Estonian resource grammar
  • Java front-end + various Python scripts for working with the GF webservice
• language skills
  • good: Estonian, English
  • bad: German, Dutch, Finnish, Hungarian, Russian

• functional programming language for grammar engineering
• parsing and generation (linearizing)
• focus on multilinguality
  • multiple concrete grammars
  • common single abstract grammar (language-neutral)
  • translate = parse string in concrete language A to abstract tree + linearize tree as a string in concrete language B
• special support for natural language features
  • long-distance dependencies
  • word form generation
  • Resource Grammar Library (RGL)
• use case: defining multilingual controlled natural languages
• developed by Aarne Ranta et al at the University of Gothenburg

String

this delicious cheese is very Italian (English)

Language-neutral tree

Strings

questo formaggio delizioso è molto italiano (Italian, notice word order)
see maitsev juust on väga itaaliapärane (Estonian)
...

• multilingual natural language processing applications
  • map between a natural language and another natural language
  • map between a natural language and a formal language
• complex NLP applications defined entirely by a grammar, where the complexity is about:
  • multilinguality
  • multimodality (touch, speech, ...)
  • dialogue
• applications with small (and controlled) vocabulary and grammar
  • software localization
  • although: recent work on parsing large corpora (Penn treebank) with GF (see: Krasimir Angelov)

• grammar-based translation (e.g. GF)
  • high precision
  • low coverage
  • suitable for translating human-computer interfaces
• statistical machine translation (e.g. Google Translate)
  • low precision
  • high coverage
  • suitable for translating arbitrary webpages
• producer vs consumer
  • when I write (produce) a letter I care about quality (e.g. use gengo.com)
  • when I read (consume) a letter I use Google Translate
• GF targets producers

Benefit: n translation descriptions, not n*n

• GF website: http://www.grammaticalframework.org/
• mailing list: https://groups.google.com/forum/?fromgroups=#!forum/gf-dev
• tutorial: http://www.grammaticalframework.org/doc/tutorial/gf-tutorial.html
• book: http://www.grammaticalframework.org/gf-book/
• summer school: http://school.grammaticalframework.org/2013/

Abstract syntax

abstract Hello = {

  flags startcat = Greeting ;

  cat Greeting ; Recipient ;

  fun
    Hello : Recipient -> Greeting ;
    World, Mum, Friends : Recipient ;
}

Concrete syntax

concrete HelloEng of Hello = {

  lincat Greeting, Recipient = {s : Str} ;

  lin
    Hello recip = {s = "hello" ++ recip.s} ;
    World = {s = "world"} ;
    Mum = {s = "mum"} ;
    Friends = {s = "friends"} ;
}

Finnish

-- This is a comment
concrete HelloFin of Hello = {
  lincat Greeting, Recipient = {s : Str} ;
  lin
    Hello recip = {s = "terve" ++ recip.s} ;
    World = {s = "maailma"} ;
    Mum = {s = "äiti"} ;
    Friends = {s = "ystävät"} ;
}

Italian

{-
   This is a multiline comment
-}
concrete HelloIta of Hello = {
  lincat Greeting, Recipient = {s : Str} ;
  lin
    Hello recip = {s = "ciao" ++ recip.s} ;
    World = {s = "mondo"} ;
    Mum = {s = "mamma"} ;
    Friends = {s = "amici"} ;
}

• abstract syntax:
  • Greeting, where we greet a Recipient, which can be World or Mum or Friends
• a comment (optional), saying what the module is doing
• a module header indicating that it is an abstract syntax module named Hello
• a module body in braces, consisting of
  • a startcat flag declaration stating that Greeting is the default start category for parsing and generation
  • category declarations introducing two categories, i.e. types of meanings
  • function declarations introducing three meaning-building functions

• a module header indicating that it is a concrete syntax named HelloEng of the abstract syntax Hello
• a module body in curly brackets, consisting of
  • linearization type definitions stating that Greeting and Recipient are records holding a single string in the field s
  • linearization definitions telling what records are assigned to each of the meanings defined in the abstract syntax
• notice:
  • concatenation: ++
  • record: { s : Str }, { s = "world" }
  • record projection: recip.s

Compile info PGF and load it

$ gf --make Hello???.gf
$ gf Hello.pgf

GF commandline examples

> parse "hello world"
Hello World

> parse "hello dad"
Unknown words: dad

> parse "world hello"
no tree found

> linearize Hello World
hello world

> parse -lang=HelloEng "hello friends" | linearize
terve ystävät
ciao amici
hello friends

Look at the internal grammar representation

> print_grammar

• multilinguality: the message is printed in many languages.
• reversibility: in addition to printing, you can parse the message and translate it to other languages.
• grammar = abstract syntax + one or more concrete syntaxes

Token sequence

• sequence of strings
• obtained by lexing an input string
  • e.g. "Hello, John!" -> ["hello", ",", "John", "!"]
• input to parser
• output of linearizer

Tree

• structure of abstract function names
  • e.g. (f1 (f2 f3))
  • other forms of trees are also possible, e.g. (f1 (f2 ?1))
• output of parser
• input to linearizer

Parse

• analyze token sequence (in a given language, assuming a category)
• ... and map it to a set of trees
  • 0 trees: parsing failed
  • 1 tree: token sequence has only one "meaning"
  • 2 or more trees: token sequence is ambiguous

Linearize

• analyze a tree
• ... and map it to a token sequence (in some language)
• variants = sequences that correspond to the same tree in the same language
  • e.g. "do not" and "don't"

Examples from MOLTO Phrasebook

• are you German?
  • 8 readings in Phrasebook
  • "you": sg/pl, man/woman, polite/informal
  • all 8 are expressed in some Romance languages
• hello!
  • 2 readings in Phrasebook
  • only revealed in Thai ("hello by man" and "hello by woman")

Example from ACE-in-GF

p -lang=Dut "John ziet precies 2 personen , die slechts reizigers inspecteren ."
| l -lang=Ace,Fin

John sees exactly 2 persons who inspect nothing but travelers .
John näkee tasan 2 henkilöä , joka tarkastaa vain matkustajia .

John sees exactly 2 persons who nothing but travelers inspect .
John näkee tasan 2 henkilöä , jonka vain matkustajat tarkastavat .

abstract Unitconv = {
  flags startcat = Unitconv ;
  cat Unit ; Unitconv ;
  fun
    f1 : Unit -> Unit -> Unitconv ;
    f2, f3: Unit ;
}

concrete UnitconvDut of Unitconv = {
  lincat Unit, Unitconv = {s : Str} ;
  lin
    f1 x y = {s = "hoeveel is" ++ x.s ++ "in" ++ y.s ++ "?"} ;
    f2 = {s = "mijl"} ;
    f3 = {s = "nautische mijl" | "mijl"} ;
}

concrete UnitconvWolfram of Unitconv = {
  lincat Unit, Unitconv = {s : Str} ;
  lin
    f1 x y = {s = "convert" ++ x.s ++ "to" ++ y.s} ;
    f2 = {s = "mile"} ;
    f3 = {s = "nmi"} ;
}

Parsing i.e. converting a string hoeveel is nautische mijl ... to tree(s)

Unitconv> parse -lang=Dut "hoeveel is nautische mijl in mijl ?"

f1 f3 f2
f1 f3 f3

Linearization i.e. converting a tree f1 f3 f2 to string(s)

Unitconv> linearize -treebank -list (f1 f3 f2)

UnitconvDut: hoeveel is nautische mijl in mijl ?, , hoeveel is mijl in mijl ?
UnitconvWolfram: convert nmi to mile

Translation i.e. parse + linearize

Unitconv> parse -lang=Dut "hoeveel is nautische mijl in mijl ?" | l -lang=Wolfram

convert nmi to mile
convert nmi to nmi

• it is up to the application designer what is considered ambiguous and synonymous
• example: synonymous or not?
  • you should perform the intake in order to publish the invitation.
  • you must perform the intake in order to publish the invitation.
  • in order to publish the invitation, you should perform the intake.
  • you should perform the intake to publish the invitation.
  • perform the intake to publish the invitation!
• example: "hello"
  • unambiguous, but ...
  • should be made ambiguous when Thai is added to the grammar

• morphological analysis: find out the possible inflection forms of words
• morphological synthesis: generate all inflection forms of words
• random generation: generate random expressions
• corpus generation: generate all expressions
• treebank generation: generate a list of trees with their linearizations
• teaching quizzes: train morphology and translation
• multilingual authoring: create a document in many languages simultaneously
• speech input: optimize a speech recognition system for a grammar

• build a larger grammar: phrases about food in English and Italian
• learn to write reusable library functions ("operations")
• learn the basics of GF's module system

abstract Food = {

  flags startcat = Phrase ;

  cat
    Phrase ; Item ; Kind ; Quality ;

  fun
    Is : Item -> Quality -> Phrase ;
    This, That : Kind -> Item ;
    QKind : Quality -> Kind -> Kind ;
    Wine, Cheese, Fish : Kind ;
    Very : Quality -> Quality ;
    Fresh, Warm, Italian, Expensive, Delicious, Boring : Quality ;
}

concrete FoodEng of Food = {

  lincat
    Phrase, Item, Kind, Quality = {s : Str} ;

  lin
    Is item quality = {s = item.s ++ "is" ++ quality.s} ;
    This kind = {s = "this" ++ kind.s} ;
    That kind = {s = "that" ++ kind.s} ;
    QKind quality kind = {s = quality.s ++ kind.s} ;
    Wine = {s = "wine"} ;
    Cheese = {s = "cheese"} ;
    Fish = {s = "fish"} ;
    Very quality = {s = "very" ++ quality.s} ;
    Fresh = {s = "fresh"} ;
    Warm = {s = "warm"} ;
    Italian = {s = "Italian"} ;
    Expensive = {s = "expensive"} ;
    Delicious = {s = "delicious" | "exquisit" | "tasty"} ; -- NB: variants
    Boring = {s = "boring"} ;
}

concrete FoodIta of Food = {

  lincat
    Phrase, Item, Kind, Quality = {s : Str} ;

  lin
    Is item quality = {s = item.s ++ "è" ++ quality.s} ;
    This kind = {s = "questo" ++ kind.s} ;
    That kind = {s = "quel" ++ kind.s} ;
    QKind quality kind = {s = kind.s ++ quality.s} ; -- NB: word order
    Wine = {s = "vino"} ;
    Cheese = {s = "formaggio"} ;
    Fish = {s = "pesce"} ;
    Very quality = {s = "molto" ++ quality.s} ;
    Fresh = {s = "fresco"} ;
    Warm = {s = "caldo"} ;
    Italian = {s = "italiano"} ;
    Expensive = {s = "caro"} ;
    Delicious = {s = "delizioso"} ;
    Boring = {s = "noioso"} ;
}

• problem in the previous examples: too much repetition
  • ++
  • {s = .. }
  • lots of similar lin structures
• solution
  • write an operator

Example

Instead:

This kind = {s = "questo" ++ kind.s} ;
That kind = {s = "quel" ++ kind.s} ;

we want to write:

This = prefix "questo" ;
That = prefix "quel" ;

resource StringOper = {
  oper
    SS : Type = {s : Str} ;
    ss : Str -> SS = \x -> {s = x} ;
    cc : SS -> SS -> SS = \x,y -> ss (x.s ++ y.s) ;
    prefix : Str -> SS -> SS = \p,x -> ss (p ++ x.s) ;
}

Explanation

• SS becomes a shorthand for {s : Str} (record with a single string)
• ss constructs a SS from a given string Str
• cc concatenates 2 SS args and returns SS
• prefix is the same as cc, but the 1st arg is a simple string

concrete FoodEng of Food = open StringOper in {
  lincat S, Item, Kind, Quality = SS ;
  lin
    Is item quality = cc item (prefix "is" quality) ;
    This = prefix "this" ; -- same as: This k = prefix "this" k ;
    That = prefix "that" ;
    QKind k q = cc k q ;
    Wine = ss "wine" ;
    Cheese = ss "cheese" ;
    Fish = ss "fish" ;
    Very = prefix "very" ;
    Fresh = ss "fresh" ;
    Warm = ss "warm" ;
    Italian = ss "Italian" ;
    Expensive = ss "expensive" ;
    Delicious = ss "delicious" ;
    Boring = ss "boring" ;
}

Notice:

• Wine takes no arguments and has lincat SS, ss "wine" produces SS
• This takes 1 argument (SS) and has lincat SS, prefix "this" also takes one argument (SS) and produces SS

Idea: for readability reasons reuse the same name for different functions.

This cannot be done directly but can be done via the overload statement.

oper mkN : overload {
  mkN : (dog : Str) -> Noun ;         -- regular nouns
  mkN : (mouse,mice : Str) -> Noun ;  -- irregular nouns
}

The definition can be also given at the same time:

oper mkN = overload {
  mkN : (dog : Str) -> Noun = regNoun ;
  mkN : (mouse,mice : Str) -> Noun = mkNoun ;
}

where regNoun is an operator that takes a string and produces Noun, and mkNoun is one that takes two strings as arguments.

abstract Morefood = Food, Fruit, Mushroom ** {
  cat
    Question ;
  fun
    QIs : Item -> Quality -> Question ;
    Pizza : Kind ;
}

concrete MorefoodIta of Morefood = FoodIta, FruitIta ** open StringOper in {
  lincat
    Question = SS ;
  lin
    QIs item quality = ss (item.s ++ "è" ++ quality.s) ;
    Pizza = ss "pizza" ;
}

• extending applies to all names (funs, lins, opers)
• by default everything is copied over
  • potential for naming conflicts!
• restricted inheritance
  • Fruit [Peach,Apple]: include only Peach and Apple from Fruit
  • Fruit - [Peach,Apple]: include everything but exclude Peach and Apple

• implement sophisticated linguistic structures:
  • morphology: the inflection of words
  • agreement: rules for selecting word forms in syntactic combinations
• cover all GF constructs for concrete syntax
• NB: the following only applies to the concrete syntax

Parameter type Number enumerating its constructors.

param Number = Sg | Pl ;

Table that depends on Number used in a linearization type.

lincat Kind = {s : Number => Str} ;

The linearization of Cheese is now a record that holds a table whose leaves are strings. This models the fact that the form of Cheese depends on Number (parametric feature).

lin Cheese = {
  s = table {
    Sg => "cheese" ;
    Pl => "cheeses"
  }
} ;

To be able to extract the string (which we need for the surface form) we use the selection operator !, e.g.

table {Sg => "cheese" ; Pl => "cheeses"} ! Pl

which returns "cheeses".

Constructors can take arguments from other parameter types.

param Number = Sg | Pl ;
param VerbForm = VPresent Number | VPast | VPastPart | VPresPart ;

A table VerbForm => Str:

table {
  VPresent Sg => "drinks" ;
  VPresent Pl => "drink" ;
  VPast       => "drank" ;
  VPastPart   => "drunk" ;
  VPresPart   => "drinking"
  }

A word is really a more complex structure than a string.

oper regNoun : Str -> {s : Number => Str} = \dog -> {
  s = table {
    Sg => dog ;
    Pl => dog + "s"
    }
  } ;

oper regVerb : Str -> {s : VerbForm => Str} = \talk -> {
  s = table {
    VPresent Sg => talk + "s" ;
    VPresent Pl => talk ;
    VPresPart   => talk + "ing" ;
    _           => talk + "ed" -- i.e. VPast (I asked), VPastPart (asked by)
    }
  } ;

Notice

• glue operator + for concatenating strings into a single token (can be applied only to strings known at compile time)
• "catch all" case _ applies to everything which the pattern matcher did not yet match
• linguistic remark: exceptions like "mice" and "seen" are not handled

Abstract definition of Is does not specify agreement restrictions:

-- this/these pizza/pizzas is/are warm
fun Is : Item -> Quality -> Phrase ;

Copula (koppelwerkwoord) depends on the number (in English):

-- copula Sg ==> "is"
oper copula : Number -> Str = \n ->
  case n of {
    Sg => "is" ;
    Pl => "are"
    } ;

Item (e.g. "this wine") should contain information about its number to be able to pass it along.

lincat Item = {s : Str ; n : Number} ;

lin This kind = {
  s = "this" ++ kind.s ! Sg ;
  n = Sg
} ;

We can now form a correct sentence

lin Is item qual = {s = item.s ++ copula item.n ++ qual.s} ;

Case expressions are syntactic sugar for tables:

case e of {...} ===  table {...} ! e

i.e. these definitions are equivalent:

oper copula : Number -> Str = \n ->
  case n of {
    Sg => "is" ;
    Pl => "are"
    } ;

oper copula : Number -> Str = \n ->
  table {
    Sg => "is" ;
    Pl => "are"
    } ! n ;

Kinds have number as a parametric feature: both singular and plural can be formed,

lincat Kind = {s : Number => Str} ;

Items have number as an inherent feature: they are inherently either singular or plural,

lincat Item = {s : Str ; n : Number} ;

Italian Kind will have parametric number and inherent gender:

lincat Kind = {s : Number => Str ; g : Gender} ;

Questions to ask when designing parameters:

• existence: what forms are possible to build by morphological and other means?
• need: what features are expected via agreement or government?

English has also cases:

param Case = Nom | Gen ;

and noun forms depend on both number and case:

oper Noun : Type = {s : Number => Case => Str} ;

We can define the worst-case function that builds the internal English noun structure like this (where x = singular form, y = plural form):

oper mkNoun : Str -> Str -> Noun = \x,y -> {
  s = table {
    Sg => table {
      Nom => x ;
      Gen => x + "'s"
      } ;
    Pl => table {
      Nom => y ;
      Gen => y + case last y of {
        "s" => "'" ;
        _   => "'s"
      }
    }
  } ;

For convenience reasons and thanks to the fact that English plural nouns usually end with "s", we can define a 1-argument version:

oper regNoun : Str -> Noun = \x -> mkNoun x (x + "s")

Note that it hides the internal noun structure, i.e. only mkNoun operates with that.

We can make the operator even smarter (more accurate):

oper regNoun : Str -> Noun = \w ->
  let
    ws : Str = case w of {
      _ + ("a" | "e" | "i" | "o") + "o" => w + "s" ;  -- bamboo
      _ + ("s" | "x" | "sh" | "o")      => w + "es" ; -- bus, hero
      _ + "z"                           => w + "zes" ;-- quiz
      _ + ("a" | "e" | "o" | "u") + "y" => w + "s" ;  -- boy
      x + "y"                           => x + "ies" ;-- fly
      _                                 => w + "s"    -- car
      }
  in
  mkNoun w ws

GF supports regular expression patterns:

_ + ("a" | "e" | "i" | "o") + "o" => w + "s" ;  -- bamboo
_ + ("s" | "x" | "sh" | "o")      => w + "es" ; -- bus, hero
x + "y"                           => x + "ies" ;-- fly
_                                 => w + "s"    -- car

Note:

• disjunctive patterns P | Q
• concatenation patterns P + Q
• variable (x) matches anything and gets bound to it
• ordering matters, e.g. the suffix "oo" prevents "bamboo" from matching the suffix "o".

Idea: operator calculates the word full paradigm from as little input information as possible. As a result the application lexicon will look simple and easy to modify.

English

Dog = mkN "dog"
Mouse = mkN "mouse" "mice"
...

German

Country = mkN "Land" neutr
...

Concise notation for tables

\\x1,...,xn => t === table { x1 => ... table { xn => t } ... }

Example

param Number = Sg | Pl ;
      Gender = Masc | Fem ;

lincat
  Quality = {s : Gender => Number => Str} ;
  Kind = {s : Number => Str ; g : Gender} ;

lin
  QKind quality kind = {
    s = table {
      Sg => kind.s ! Sg ++ quality.s ! kind.g ! Sg ;
      Pl => kind.s ! Pl ++ quality.s ! kind.g ! Pl ;
    };
    g = kind.g
  } ;

  -- NB: shorter
  QKind quality kind = {
    s = \\n => kind.s ! n ++ quality.s ! kind.g ! n ;
    g = kind.g
  } ;

param
  Number = Sg | Pl ;
  Gender = Masc | Fem ;

lincat
  Quality = {s : Gender => Number => Str} ;
  Kind = {s : Number => Str ; g : Gender} ;

lin
  Very qual = {s = table {
        Masc => table {
            Sg => "molto" + qual.s ! Masc ! Sg ;
            Pl => "molto" + qual.s ! Masc ! Pl
        } ;
        Fem => table {
            Sg => "molto" + qual.s ! Fem ! Sg ;
            Pl => "molto" + qual.s ! Fem ! Pl
        }
    }
  }

  -- NB: shorter
  Very qual = {s = \\g,n => "molto" ++ qual.s ! g ! n} ;

Example: German transitive verbs (i.e. verbs that syntactically require an object) determine the case of their object (i.e. this is an inherent feature).

Clean solution: extend the regular verb type and definitions:

lincat TV = Verb ** {c : Case} ;
lin Follow = regVerb "folgen" ** {c = Dative} ;

Benefits

• TV can be used in contexts where Verb is required

-- Optional string with preference on the string vs. empty.
optStr : Str -> Str = \s -> variants {s ; []} ;
strOpt : Str -> Str = \s -> variants {[] ; s} ;

-- Infix
infixSS   : Str  -> SS -> SS -> SS = \f,x,y -> ss (x.s ++ f ++ y.s) ;
prefixSS  : Str        -> SS -> SS = \f,x   -> ss (f ++ x.s) ;
...

-- Bind together two tokens in some lexers, either obligatorily or optionally
glue : Str -> Str -> Str = \x,y -> x ++ BIND ++ y ;
glueOpt : Str -> Str -> Str = \x,y -> variants {glue x y ; x ++ y} ;
noglueOpt : Str -> Str -> Str = \x,y -> variants {x ++ y ; glue x y} ;

-- Force capitalization of next word in some unlexers
capitalize : Str -> Str = \s -> CAPIT ++ s ;

-- These should be hidden, and never changed since they are hardcoded
-- in (un)lexers
BIND : Str = "&+" ;
PARA : Str = "&-" ;
CAPIT : Str = "&|" ;

-- Parentheses
paren : Str -> Str = \s -> "(" ++ s ++ ")" ;
parenss : SS -> SS = \s -> ss (paren s.s) ;

-- Missing form.
nonExist : Str = variants {} ;

-- Identity function
id : (A : Type) -> A -> A = \_,a -> a ;

-- Zero, one, two, or more (elements in a list etc)
param ENumber = E0 | E1 | E2 | Emore ;

eNext : ENumber -> ENumber = \e -> case e of {
  E0 => E1 ; E1 => E2 ; _ => Emore
} ;

-- Use the glue operator with arbitrary number of arguments
-- (where arbitrary =< 5 ;))
BIND : Str = "&+" ;
glue = overload {
  glue : (x1,x2 : Str) -> Str = \x1,x2 -> x1 ++ BIND ++ x2 ;
  glue : (x1,x2,x3 : Str) -> Str = \x1,x2,x3 -> x1 ++ BIND ++ x2 ++ BIND ++ x3;
  glue : (x1,x2,x3,x4 : Str) -> Str = \x1,x2,x3,x4 -> x1 ++ BIND ++ x2 ++ ... 
  glue : (x1,x2,x3,x4,x5 : Str) -> Str = \x1,x2,x3,x4,x5 -> x1 ++ BIND ++ ... 
};

param Bool = True | False ;

oper

if_then_else : (A : Type) -> Bool -> A -> A -> A = \_,c,d,e ->
  case c of {
    True => d ;
    False => e
  } ;

andB : (_,_ : Bool) -> Bool = \a,b -> if_then_else Bool a b False ;
orB  : (_,_ : Bool) -> Bool = \a,b -> if_then_else Bool a True b ;
notB : Bool         -> Bool = \a   -> if_then_else Bool a False True ;

if_then_Str : Bool -> Str -> Str -> Str = if_then_else Str ;

onlyIf : Bool -> Str -> Str = \b,s -> case b of {
  True => s ;
  _ => nonExist
} ;

• contain the linguistic knowledge
• provide a language-neutral API over many languages
• currently 26 languages
  • mostly European languages

• Synopsis: http://www.grammaticalframework.org/lib/doc/synopsis.html
• coverage: http://www.postcrashgames.com/gf_world/index.html

A resource grammar has two kinds of categories and two kinds of rules:

• lexical:
  • lexical categories, to classify words
  • lexical rules, to define words and their properties
• phrasal (combinatorial, syntactic):
  • phrasal categories, to classify phrases of arbitrary size
  • phrasal rules, to combine phrases into larger phrases

GF makes no formal distinction between these two kinds. But it is a good discipline to follow.

Two kinds of lexical categories:

• closed:
  • a finite number of words
  • seldom extended in the history of language
  • structural words / function words

Example:

  Conj ;     -- conjunction           e.g. "and"
  Det ;      -- determiner            e.g. "this"

• open:
  • new words are added all the time
  • content words

Example:

  N ;        -- noun         e.g. "pizza"
  A ;        -- adjective    e.g. "good"
  V ;        -- verb         e.g. "sleep"

this_Det, that_Det, these_Det, those_Det : Det ;
very_AdA  : AdA ;

Cl ;   -- clause             e.g. "this pizza is good"
NP ;   -- noun phrase        e.g. "this pizza"
CN ;   -- common noun        e.g. "warm pizza"
AP ;   -- adjectival phrase  e.g. "very warm"
...

We need the following combinations:

mkCl : NP -> AP -> Cl ;      -- e.g. "this pizza is very warm"
mkNP : Det -> CN -> NP ;     -- e.g. "this pizza"
mkCN : AP -> CN -> CN ;      -- e.g. "warm pizza"
mkAP : AdA -> AP -> AP ;     -- e.g. "very warm"

We also need lexical insertion, to form phrases from single words:

mkCN : N -> NP ;
mkAP : A -> AP ;

Naming convention: to construct a C, use a function mkC.

Heavy overloading: e.g. ~20 operations named mkNP!

Language-specific and language-independent parts:

• the syntax API Syntax??? has the same types and functions for all languages
• the morphology API Paradigms??? has partly different types and functions for different languages

Dutch

> i -path=alltenses -retain alltenses/ParadigmsDut.gfo
> cc -table mkN "auto"
s . ResDut.NF ParamX.Sg ResDut.Nom => auto
s . ResDut.NF ParamX.Sg ResDut.Gen => autos
s . ResDut.NF ParamX.Pl ResDut.Nom => auto's
s . ResDut.NF ParamX.Pl ResDut.Gen => auto's
g . ResDut.Utr

Finnish

> i -path=alltenses -retain alltenses/ParadigmsFin.gfo
> cc -table mkN "talo"
s . ResFin.NCase ParamX.Sg ResFin.Nom => talo
s . ResFin.NCase ParamX.Sg ResFin.Gen => talon
s . ResFin.NCase ParamX.Sg ResFin.Part => taloa
s . ResFin.NCase ParamX.Sg ResFin.Transl => taloksi
...
...
...

lincat
  Phrase = Cl ;
  Item = NP ;
  Kind = CN ;
  Quality = AP ;

lin
  Is item quality = mkCl item quality ;
  This kind = mkNP this_Det kind ;
  That kind = mkNP that_Det kind ;
  These kind = mkNP these_Det kind ;
  Those kind = mkNP those_Det kind ;
  QKind quality kind = mkCN quality kind ;
  Very quality = mkAP very_AdA quality ;

Lexical rules (English-specific):

  Wine = mkCN (mkN "wine") ;
  Pizza = mkCN (mkN "pizza") ;
  Cheese = mkCN (mkN "cheese") ;
  -- NB: 'fish' is irregular word, 1-arg smart paradigm would fail
  Fish = mkCN (mkN "fish" "fish") ;
  Fresh = mkAP (mkA "fresh") ;
  Warm = mkAP (mkA "warm") ;
  Italian = mkAP (mkA "Italian") ;
  Expensive = mkAP (mkA "expensive") ;

• functor is a module that opens one or more interfaces
• interface is a module similar to resource, but it only contains the types of opers, not (necessarily) their definitions

Add the keyword incomplete. We will use the header

incomplete concrete FoodsI of Foods = open Syntax, LexFoods in

where

interface Syntax    -- the resource grammar interface
interface LexFoods  -- the domain lexicon interface

When we moreover have

instance SyntaxEng of Syntax     -- the English resource grammar
instance LexFoodsEng of LexFoods -- the English domain lexicon

we can write a functor instantiation

concrete FoodsGer of Foods = FoodsI with
  (Syntax = SyntaxGer),
  (LexFoods = LexFoodsGer) ;

incomplete concrete FoodsI of Foods = open Syntax, LexFoods in {
lincat
  Phrase = Cl ;
  Item = NP ;
  Kind = CN ;
  Quality = AP ;
lin
  Is item quality = mkCl item quality ;
  This kind = mkNP this_Det kind ;
  That kind = mkNP that_Det kind ;
  These kind = mkNP these_Det kind ;
  Those kind = mkNP those_Det kind ;
  QKind quality kind = mkCN quality kind ;
  Very quality = mkAP very_AdA quality ;

  Wine = mkCN wine_N ;
  Pizza = mkCN pizza_N ;
  Cheese = mkCN cheese_N ;
  Fish = mkCN fish_N ;
  Fresh = mkAP fresh_A ;
  Warm = mkAP warm_A ;
  Italian = mkAP italian_A ;
  ...
}

interface LexFoods = open Syntax in {
oper
  wine_N : N ;
  pizza_N : N ;
  cheese_N : N ;
  fish_N : N ;
  fresh_A : A ;
  warm_A : A ;
  italian_A : A ;
  expensive_A : A ;
  delicious_A : A ;
  boring_A : A ;
}

Definitions for the opers in the instance.

instance LexFoodsGer of LexFoods = open SyntaxGer, ParadigmsGer in {
oper
  wine_N = mkN "Wein" ;
  pizza_N = mkN "Pizza" "Pizzen" feminine ;
  cheese_N = mkN "Käse" "Käsen" masculine ;
  fish_N = mkN "Fisch" ;
  fresh_A = mkA "frisch" ;
  warm_A = mkA "warm" "wärmer" "wärmste" ;
  italian_A = mkA "italienisch" ;
  expensive_A = mkA "teuer" ;
  delicious_A = mkA "köstlich" ;
  boring_A = mkA "langweilig" ;
}

Super simple! Just parametrizes the functor (FoodsI) with the German grammar and the domain lexicon.

concrete FoodsGer of Foods = FoodsI with
  (Syntax = SyntaxGer),
  (LexFoods = LexFoodsGer) ;

Just two modules are needed:

• a domain lexicon instance
• a functor instantiation

The functor instantiation is completely mechanical to write.

The domain lexicon instance requires some knowledge of the words of the language:

• what words are used for which concepts
• how the words are constructed
• features such as gender

Module types:

• abstract: Foods
• concrete: FoodsFin, FoodsI (incomplete)
• interface: Syntax, LexFoods
• instance: SyntaxFin, LexFoodsFin

AdjectiveDut.gf
AdverbDut.gf
AllDutAbs.gf
AllDut.gf
CatDut.gf
ConjunctionDut.gf
ExtDut.gf
ExtraDutAbs.gf
ExtraDut.gf
GrammarDut.gf
IdiomDut.gf
IrregDutAbs.gf
IrregDut.gf
LangDut.gf
LexiconDut.gf
MakeStructuralDut.gf: resource
MorphoDut.gf: resource
NounDut.gf
NumeralDut.gf
ParadigmsDut.gf: resource
PhraseDut.gf
QuestionDut.gf
RelativeDut.gf
ResDut.gf
SentenceDut.gf
StructuralDut.gf
SymbolDut.gf
VerbDut.gf

Adjective.gf
Adverb.gf
Backward.gf
Cat.gf
Common.gf
Compatibility.gf
Conjunction.gf
Extra.gf
Grammar.gf
Idiom.gf
Lang.gf
Lexicon.gf
Noun.gf
Numeral.gf
NumeralTransfer.gf
Phrase.gf
Question.gf
Relative.gf
Sentence.gf
Structural.gf
Symbol.gf
Tense.gf
Text.gf
Transfer.gf
Verb.gf

Combinators.gf: incomplete resource
CombinatorsDut.gf: resource
...
Constructors.gf: incomplete resource
ConstructorsDut.gf: resource
...
Syntax.gf: interface
SyntaxDut.gf: instance
...
Symbolic.gf: incomplete resource
SymbolicDut.gf: resource
...
TryDut.gf
...

• language-neutral
• categories (cat) and functions (fun)
• at most one startcat (can be also given at runtime)
• each function has exactly one type
  • type is a category
  • or a structure of categories
• dependent types (see Topic 5)

abstract Test = {
    flags startcat = Greeting ;
    cat Greeting ;
    cat Name ;

    fun hi : Greeting ;
    fun personal_hi : Name -> Greeting ;
    fun two_person_hi : Name -> Name -> Greeting ;

    -- Note: you cannot use the same function name
    -- with a different type
    -- fun hi : Name -> Greeting ;

    -- this is OK because -> is right assoc.
    fun two_person_hi : Name -> (Name -> Greeting) ;

    fun weird_hi : (Name -> Name) -> Greeting ;
}

• concrete language (e.g. Dutch)
• of some abstract module (e.g. some application)
• category linearization type (lincat)
• function linearization (lin)
• possibly complex linearization types
  • string
  • record
  • table
  • record of tables of tables of strings
  • ...
• advice: use records (e.g. {s : Str }) instead of strings (Str), because records can be often extended without breaking code (e.g {s : Str, g : Gender})

• can be opened
• oper definitions

• interface is a module similar to resource
• only contains the types of opers, not (necessarily) their definitions

• implements interface
• fills in the operator definitions

• anything can be declared incomplete, e.g.
  • concrete is complete with respect to an abstract
  • resource is complete if all opers and params have a definition part
• incomplete concrete can be parametrized in a concrete

• in the concrete syntax it is possible to restrict which phrases go together e.g. a noun phrase and a verb phrase must agree in number
  • 2 men run
  • 1 man runs
  • * 2 men runs
  • * 1 man run
• sometimes restrictions are language-independent, e.g.
  • convert 3 kg to grams
  • * convert 3 kg to EUR
  • dim the light
  • * dim the fan
• language-independent restrictions should be defined in the abstract syntax
• solution: dependent types

• there are commands and device kinds
• for each kind of device, there are devices and actions
• a command concerns an action of some kind on a device of the same kind

Modeling:

• Command and Kind are GF categories
• Device and Action are dependent types (they depend on Kind)
• Command does not have a Kind

Abstract

cat
  Command ;
  Kind ;
  Device Kind ; -- argument type Kind
  Action Kind ;

fun
  -- the Kinds of Action and Device must be the same,
  -- to be able to form a Command
  CAction : (k : Kind) -> Action k -> Device k -> Command ;
  DKindOne : (k : Kind) -> Device k ;
  light, fan : Kind ;
  dim : Action light ;

Concrete

The concrete syntax does not know anything about dependent types but must suppress the extra argument.

lincat Action = {s : Str} ;
-- the Kind argument is suppressed in linearization
lin CAction _ act dev = {s = act.s ++ dev.s} ;

Syntax and semantics are OK:

> parse "dim the light"
CAction light dim (DKindOne light)

Syntax OK, semantics not:

> parse "dim the fan"
The parsing is successful but the type checking failed with error(s):
  Couldn't match expected type Device light
         against inferred type Device fan
  In the expression: DKindOne fan

Token look-ahead (unfortunately) ignores dependent types:

> parse "dim the
fan    light

Sometimes an action can be performed on all kinds of devices.

This is represented as a function that takes a Kind as an argument and produces an Action for that Kind:

fun switchOn, switchOff : (k : Kind) -> Action k ;

Functions of this kind are called polymorphic.

Compare:

light : Kind ;
dim : Action light ;

Parsing examples:

Test> parse "switch on the fan "
CAction fan (switchOn fan) (DKindOne fan)

Test> parse "switch on the light "
CAction light (switchOn light) (DKindOne light)

Abstract

flags startcat = Command ;

cat Command ; Kind ; Device Kind ; Action Kind ;

fun
    CAction : (k : Kind) -> Action k -> Device k -> Command ;
    light, fan : Kind ;
    dim : Action light ;
    DKindOne : (k : Kind) -> Device k ;
    switchOn, switchOff : (k : Kind) -> Action k ;

Concrete

lincat Command, Kind, Device, Action = { s : Str };

lin CAction _ act dev = {s = act.s ++ dev.s} ;

    -- Kinds
    light = { s = "light" } ;
    fan = { s = "fan" } ;
    DKindOne k = { s = "the" ++ k.s } ;

    -- Actions
    dim = { s = "dim" } ;
    switchOn _ = { s = "switch on" } ;
    switchOff _ = { s = "switch off" } ;

• concrete syntax is unaware of dependent types
• GF's built-in token look-ahead ignores dependent types
• various other GF tools do not support dep. types

• the parser and linearizer work with tokens
• lexing (tokenization) = mapping a string to tokens
  • e.g. split punctuation from the words
• unlexing = mapping tokens to strings
  • e.g. glue punctuation to preceding words
• lexing/unlexing and application/language dependent (i.e. not general solution)
• lexing/unlexing can be done by an external application as long as some GF conventions are observed
  • space marks token border
  • &+ glues tokens

Example (token list that as a string should look like "12 is a number."):

1 &+ 2 is a number &+ .

• (see the commandline examples on the previous slides)
• can be also run in "server mode"

PGF server usage example

$ GF_RESTRICTED=yes gf --server=1234 --document-root /path/docroot/

$ (create /path/docroot/grammars/Test.pgf)

$ curl "http://localhost:1234/grammars/Test.pgf?
    command=parse&cat=Utt&from=TestEng&input=hello"

$ (process the JSON output)

• parse
  • in: string, language
  • out: tree set
• complete
  • in: string, language
  • out: token set
• linearize
  • in: tree
  • out: language -> string (+ variants?)
• translate (= parse + linearize)
• info
  • out: lists of functions, categories, startcat, ...

• Syntax editor (really a tree editor)
  • edit a tree: add/remove node, replace node with a random structure of the same category, wrap node
  • view linearizations (of a possibly partial tree)
  • tree cannot be ambiguous but linearizations can be
• Minibar
  • look-ahead edit a sentence
  • view its trees and translations
• the editors are online and integrated
  • switch to Syntax editor to edit a small bit in the beginning of an existing sentence
  • switch to Minibar to view linearization ambiguity
• links
  • http://cloud.grammaticalframework.org/minibar/minibar.html
  • http://cloud.grammaticalframework.org/syntax-editor/editor.html

• features
  • syntax highlighting and error detection
  • code folding, quick block-commenting, automatic code formatting
  • definition outlining, jump to declaration, find usage
  • warnings for problems in module dependency hierarchy
  • launch configurations, i.e. compilation directly from IDE
  • use GF Shell from within Eclipse
  • auto-completion for declared identifiers
  • background compilation (shallow) using project builder
  • support for Open Declaration (F3), including qualified names
  • code generation for new languages in application grammars
  • inline documentation for function calls, overloads
  • proper cross-reference handling with qualified names
  • test management and testing tool
  • external library browser
• http://www.grammaticalframework.org/eclipse/index.html

• http://cloud.grammaticalframework.org/gfse/
• GF syntax aware text editor
• web-based: an easy way to explore GF
• collaborative editor: grammars can be shared
• some limitations e.g. simpler grammars

• access to PGF services
  • Haskell API
  • GF Webservice (REST)
  • JPGF (Java)
  • Python API (work in progress ?)
  • C API (work in progress ?)
  • Javascript API (?)
  • Java API to GF Webservice
• GF Cloud Service beta
  • extension of GF (PGF) Webservice
  • added: grammar upload, download, recompile
  • http://cloud.grammaticalframework.org/gf-cloud-api.html

• multilingual applications (with high quality language)
  • tourist phrasebook
  • museum knowledge base
• multimodal interfaces (with different modalities handled in the same framework)
• dialog systems
• software localization (?)
• speech recognition grammars
• multilingual wiki

• basic sentences that a tourist needs in many different languages
• http://www.grammaticalframework.org/demos/phrasebook/
• portable
  • PhraseDroid: org.grammaticalframework.android.apps.phrasedroid @ Google Play

• multimodal
  • speech input
  • touch/click input
  • ...
• multilingual
  • English
  • Spanish
  • ...
• dialog-based
  • speech output
  • meaning emerges from a sequence of smaller conversation turns
• demo video: http://www.youtube.com/watch?v=1bfaYHWS6zU
• see also GF Book 7.15

• interpret an audio signal as string in language A and translate it to language B for evaluation
  • lang A: natural language, audio signal interpreted using acoustic models and a speech recognition grammar (e.g. JSGF, FSA)
  • lang B: formal language (calculator, directions query, ...)
• GF provides
  • translation (approximation): GF -> JSGF/SLF/...
  • translation: lang A -> lang B
  • modular grammar, multiple languages, smart paradigms, ...
• a smartphone app
  • records audio
  • sends it for transcription+translation
  • sends the result for evaluation/execution

Example

audio -> natural language string -> tree -> formal expression -> evaluation result

/one plus two/ -> "one plus two" -> (plus n1 n2) -> eval(1+2) -> 3

Links

• GF grammars for Estonian, English and "App"
  • http://kaljurand.github.com/Grammars/
• Android apps
  • http://play.google.com/store/apps/details?id=ee.ioc.phon.android.speak
  • http://play.google.com/store/apps/details?id=ee.ioc.phon.android.arvutaja

• wiki
  • user-friendly collaborative environment for knowledge management
  • content typically unconstrained natural language (NL)
  • powered by software, e.g. MediaWiki
  • e.g. Wikipedia
• semantic wiki (= wiki + formal semantics)
  • provides: richer query language, consistency checking (via automatic reasoning)
  • content typically NL + typed links (i.e. RDF triples)
  • software: Semantic Mediawiki, ...
• CNL-based semantic wiki (= semantic wiki using CNL)
  • formal languages hidden (=> can use more expressive formal languages)
  • software: AceWiki
• multilingual wiki
  • authoring in multiple (natural) languages
  • current systems: only document-level interlinking

• multiple languages
  • natural: English, German, ACE, ...
  • formal: ACE, FOL, ...
  • languages for content, UI, meta information
• content
  • viewable/editable/queryable in multiple languages
  • automatically kept in sync
• CNL-based
  • backed by formal grammar(s)
  • formal languages are hidden
• semantic
  • consistency checking, question answering, ...
  • precise translation

• multilingual ACE wiki
  • authoring in multiple ACE-based CNLs
• tourist phrasebook
  • book structure (ToC, chapters, index)
  • multiple languages
  • grammar editing
• catalog of museum objects (paintings, painters)
  • each object on a separate wiki page
  • multiple languages
  • rich queries (e.g. "which Dutch painter painted which French painter?")
• logic/math puzzles
  • multiple user solutions
  • automatically checked

• AceWiki
  • collaborative environment
  • GUI (e.g. look-ahead editor)
  • storage
  • connection to ACE parser
  • connection to OWL reasoners
• Grammatical Framework (GF)
  • multilingual grammars
  • parser (translation, completion, ...)
  • grammar editor

• goal: user-friendly yet expressive semantic wiki system
• wiki features: collaborative editing, multiple interlinked articles
• background reasoning language: OWL
  • expressive fragment of first-order logic
  • decidable reasoning tasks: consistency checking, question answering, ...
  • complex syntax
• front-end language: ACE
  • subset of natural English
  • well-defined translation into first-order logic / OWL
  • end-user documentation: construction and interpretation rules
• developed by Tobias Kuhn
• see more: http://attempto.ifi.uzh.ch/acewiki/

• multilingual viewing and editing of wiki content
  • grammar-based editing (show next possible tokens)
• wiki entry is GF abstract tree set
  • viewed via linearization(s)
  • can represent ambiguity
• access to multiple online GF grammars
  • provided by GF Webservice
  • single grammar per wiki
• grammar integrated into the wiki
  • wiki-linking of grammar and content
  • grammar can be updated while building the wiki
• multilingual ACE grammar implemented in GF
  • other GF grammars can be used instead (no ACE-based reasoning in this case)

• goal: user-friendly language for formal knowledge engineering
• subset of natural English
• translatable into Discourse Representation Structures (DRS)
  • and further into standard first-order logic, OWL, various rule languages
  • enables automatic reasoning, e.g. consistency checking, question answering, ...
• verbalization of formal languages
  • DRS, OWL
• end-user documentation: construction and interpretation rules
• editing environments: AceWiki, ACE Editor, ACE View, ...

An ACE grammar in GF/RGL adds multiple natural languages as front-ends to ACE.

• implementation of the ACE syntax (i.e. no DRS generation)
  • extension of Angelov and Ranta (CNL 2009)
• available in 15 natural languages via the RGL
  • Catalan, Danish, Dutch, English, Finnish, French, German, Italian, Latvian, Norwegian, Polish, Romanian, Russian, Spanish, Swedish
  • design allows for easy extendability
• status
  • focus on the subset of ACE that is used in AceWiki (almost 100% coverage at almost 0% ambiguity)
  • some precision problems, e.g. anaphoric references do not obey DRS accessibility constraints
  • ambiguity and coverage problems in some languages

p -lang=Ace "if a person admires no golfer then the person buys
    at least 2 aquariums that nothing but travelers inspect ." | l

si una persona no admira cap golfista llavors la persona compra
    almenys 2 aquarins que nomÈs viatgers inspeccionen .

als een persoon geen golfer bewondert , dan koopt de persoon
    ten minste 2 aquaria die slechts reizigers inspecteren .

jos henkilö ei ihaile mitään golfaajaa niin henkilö ostaa
    vähintään 2 akvaariota jonka vain matkustajat tarkastavat .

si une personne n' admire aucun golfeur alors la personne achète
    au moins 2 aquariums que seulement des voyageurs inspectent .

wenn eine Person keinen Golfer bewundert , dann kauft die Person
    wenigstens 2 Aquariume die nur Reisenden inspizieren .

si una persona non ammira nessuno giocatore di golf allora la persona compra
    almeno 2 acquari che soltanto viaggiatori ispezionano .

si una persona no admira hacia golfista entonces la persona compra
    al menos 2 acuarios que solamente viajeros inspeccionan .

om en person beundrar inget golfspelare så personen köper
    minst 2 akvariumar som bara resenärar avsynar .

• AceWiki-GF
  • GitHub repository: http://github.com/AceWiki (see the gfservice-branch)
  • demo wikis: http://attempto.ifi.uzh.ch/acewiki-gf/
• Multilingual ACE (ACE-in-GF)
  • http://github.com/Attempto/ACE-in-GF
• Attempto project
  • http://attempto.ifi.uzh.ch