Type Description Language and other aspects of DELPH-IN Joint Reference Formalism

TDL File Syntax

Productions are separated into thematic sections. ALL-CAPS rule names are for non-content terminals, which appear at the bottom of the description.

   1 # File Contents
   2 #
   3 # Note: The LKB does not parse environments (:begin ... :end), nor does it
   4 #       support :include statements, so the following is only applicable for
   5 #       PET, ACE, and perhaps agree.
   6 
   7 TdlFile      := ( Environment | Statement | Spacing )* EOF
   8 Environment  := BEGIN TYPE DOT TypeEnv END TYPE DOT
   9               | BEGIN INSTANCE Status? DOT InstanceEnv END INSTANCE DOT
  10 TypeEnv      := ( Environment | Statement
  11                 | TypeDef | TypeAddendum | Spacing )*
  12 InstanceEnv  := ( Environment | Statement
  13                 | InstanceDef | LetterSet | WildCard | Spacing )*
  14 Status       := STATUS ( "generic-lex-entry"
  15                        | "lex-entry"
  16                        | "lexical-filtering-rule"
  17                        | "lex-rule"
  18                        | "post-generation-mapping-rule"
  19                        | "rule"
  20                        | "token-mapping-rule" ) Spacing
  21 
  22 # Note: The LKB has several Lisp functions which open files in specified
  23 #       environments, so the following are parsing targets for those
  24 #       functions.
  25 
  26 TdlTypeFile  := ( TypeDef | TypeAddendum | Spacing )* EOF
  27 TdlRuleFile  := ( InstanceDef | LetterSet | WildCard | Spacing )* EOF
  28 
  29 # Note: Krieger & Schaeffer 1994 define a large number of statements, but
  30 #       DELPH-IN grammars appear to only use :include.
  31 # Note: :include's string argument is a path relative to the current file's
  32 #       directory. If the filename extension is not given, the default ".tdl"
  33 #       extension is used. The file is opened in the same environment as the
  34 #       :include statements (e.g., :include in a type environment opens the
  35 #       file and parses it as TypeEnv)
  36 
  37 Statement    := Include
  38 Include      := INCLUDE Filename DOT
  39 Filename     := DQString
  40 
  41 # Types and Instances
  42 #
  43 # Note: Instances may be syntactically identical to type definitions, but they
  44 #       do not affect the type hierarchy. They may also be lexical rule
  45 #       definitions that include an affixing pattern to a definition.
  46 
  47 TypeDef      := TypeName DEFOP TypeDefBody DOT
  48 TypeAddendum := TypeName ADDOP AddendumBody DOT
  49 TypeName     := Identifier Spacing
  50 
  51 InstanceDef  := TypeDef | LexRuleDef
  52 LexRuleDef   := LexRuleId DEFOP Affix? TypeDefBody DOT
  53 LexRuleId    := Identifier Spacing
  54 
  55 # Identifiers are used in several patterns
  56 #
  57 # Note: The characters disallowed in Identifiers are chosen to avoid ambiguity
  58 #       with other parts of the TDL syntax.
  59 
  60 Identifier   := /[^\s!"#$%&'(),.\/:;<=>[\]^|]+/
  61 
  62 # Definition Bodies (top-level conjunctions of terms)
  63 #
  64 # Note: Definition bodies are most simply Conjunctions, but several
  65 #       variations require special productions:
  66 #
  67 #       (1) """DocStrings""" may precede any top-level Term or the final DOT
  68 #       (2) TypeDef and LexRuleDef require at least one TypeName
  69 #       (3) TypeAddendum may use a DocString in place of a Conjunction
  70 #           
  71 
  72 TypeDefBody  := TypedConj DocString?
  73 AddendumBody := DocConj DocString? | DocString
  74 
  75 # Note: To accommodate TypeDefBody and AddendumBody, three special
  76 #       conjunctions are added:
  77 #
  78 #       (1) TypedConj has an obligatory TypeName term
  79 #       (2) FeatureConj excludes type terms (including strings, etc.)
  80 #       (3) DocConj is a regular conjunction with optional DocStrings
  81 #
  82 #       Note that FeatureConj is only necessary to reduce ambiguity (e.g.,
  83 #       for LALR parsing); if ambiguity is allowed, DocConj may be used.
  84 
  85 TypedConj    := ( FeatureConj AND )? DocString? TypeName ( AND DocConj )?
  86 FeatureConj  := DocString? FeatureTerm ( AND DocString? FeatureTerm )*
  87 DocConj      := DocString? Term ( AND DocString? Term )*
  88 
  89 # Note: The DocString pattern may span multiple lines
  90 
  91 DocString    := /"""([^"\\]|\\.|"(?!")|""(?!"))*"""/ Spacing
  92 
  93 # Terms and Conjunctions
  94 
  95 Conjunction  := Term ( AND Term )*
  96 Term         := TypeTerm | FeatureTerm | Coreference
  97 TypeTerm     := TypeName
  98               | DQString
  99               | Regex
 100 FeatureTerm  := Avm
 101               | DiffList
 102               | ConsList
 103 
 104 DQString     := ( /""(?!")/ | /"([^"\\]|\\.)+"/ ) Spacing
 105 Regex        := "^" /([^$\\]|\\.)*/ "$" Spacing
 106 
 107 Avm          := AVMOPEN AttrVals? AVMCLOSE
 108 AttrVals     := AttrVal ( COMMA AttrVal )*
 109 AttrVal      := AttrPath SPACE Conjunction
 110 AttrPath     := Attribute ( DOT Attribute )*
 111 Attribute    := Identifier Spacing
 112 
 113 DiffList     := DLOPEN Conjunctions? DLCLOSE
 114 ConsList     := CLOPEN ( Conjunctions ConsEnd? )? CLCLOSE
 115 ConsEnd      := COMMA ELLIPSIS | DOT Conjunction
 116 Conjunctions := Conjunction ( COMMA Conjunction )*
 117 
 118 Coreference  := "#" Identifier Spacing
 119 
 120 # Letter-sets, Wild-cards, and Affixes
 121 #
 122 # Note: spacing is sensitive within these patterns, so many non-content
 123 #       terminals are used directly with an explicit SPACE instead of in
 124 #       a production with Spacing.
 125 
 126 LetterSet    := "%(letter-set" SPACE? LetterSetDef SPACE? ")"
 127 WildCard     := "%(wild-card" SPACE? WildCardDef SPACE? ")"
 128 LetterSetDef := "(" LetterSetVar SPACE Characters ")"
 129 WildCardDef  := "(" WildCardVar SPACE Characters ")"
 130 LetterSetVar := /![^ ]/
 131 WildCardVar  := /\?[^ ]/
 132 Characters   := /([^)\\]|\\.)+/
 133 
 134 # Note: When a LetterSetVar is used in an AffixMatch, the same LetterSetVar
 135 #       in the AffixSub copies the matched character, in order, so there
 136 #       should be the same number of LetterSetVars in both, but this is not
 137 #       captured in the syntax.
 138 
 139 Affix        := AffixClass AffixPattern+ Spacing
 140 AffixClass   := "%prefix" | "%suffix"
 141 AffixPattern := SPACE? "(" AffixMatch SPACE AffixSub ")"
 142 AffixMatch   := NullChar | CharList
 143 AffixSub     := CharList
 144 NullChar     := "*"
 145 CharList     := ( LetterSetVar | WildCardVar | AffixChar )+
 146 AffixChar    := /([^!?\s*\\]|\\[^ ])+/
 147 
 148 # Whitespace and Comments
 149 #
 150 # Note: SPACE and BlockComment may span multiple lines. Also, while block
 151 #       comments in Lisp may be nested (`#| outer #| inner |# outer |#`),
 152 #       support for nested comments in TDL is mixed (ACE supports it, the
 153 #       LKB does not), so this definition does not nest.
 154 
 155 Spacing      := SPACE? Comment*
 156 SPACE        := /\s+/
 157 Comment      := ( LineComment | BlockComment ) SPACE?
 158 LineComment  := /;.*$/
 159 BlockComment := "#|" /([^|\\]|\\.|\|(?!#))*/ "|#"
 160 
 161 # Non-content Terminals
 162 
 163 BEGIN        := ":begin" Spacing
 164 TYPE         := ":type" Spacing
 165 INSTANCE     := ":instance" Spacing
 166 STATUS       := ":status" Spacing
 167 INCLUDE      := ":include" Spacing
 168 END          := ":end" Spacing
 169 DEFOP        := ":=" Spacing
 170 ADDOP        := ":+" Spacing
 171 DOT          := "." Spacing
 172 AND          := "&" Spacing
 173 COMMA        := "," Spacing
 174 AVMOPEN      := "[" Spacing
 175 AVMCLOSE     := "]" Spacing
 176 DLOPEN       := "<!" Spacing
 177 DLCLOSE      := "!>" Spacing
 178 CLOPEN       := "<" Spacing
 179 CLCLOSE      := ">" Spacing
 180 ELLIPSIS     := "..." Spacing
 181 EOF          := ""  # end-of-file

Deprecated TDL Features

The following are deprecated features of DELPH-IN TDL. They are no longer considered part of the format, but implementers of TDL parsers may want to include them for backward compatibility. If so, they are encouraged to print warnings upon encountering the deprecated forms so grammar developers know to change them.

Subtyping Operator (:<)

The :< operator was originally used only for declaring a type's position in the type hierarchy (i.e., features could not be specified, unlike with :=), but eventually this constraint was relaxed and it became equivalent to :=. As of Autumn 2018, the form has been removed and is no longer considered part of DELPH-IN TDL, but the change to TDL syntax to support the operator is minimal:

   1 DEFOP        := ( ":=" | ":<" ) Spacing

Single-quoted Symbols ('symbol)

Double-quoted strings and identifiers are both type names, but there used to be Lisp-like single-quoted symbols as well. These still exist in some grammars, such as those using an old version of matrix.tdl, which has the following:

implicit-coord-rel := coordination-relation &
  [ PRED 'implicit_coord_rel ].
null-coord-rel := coordination-relation &
  [ PRED 'null_coord_rel ].

There is no difference between using quoted symbols and regular strings or identifiers (although identifiers would need to be defined as types somewhere), so recent versions of matrix.tdl have this instead:

implicit-coord-rel := coordination-relation &
  [ PRED "implicit_coord_rel" ].
null-coord-rel := coordination-relation &
  [ PRED "null_coord_rel" ].

The change to the syntax to support quoted symbols is as follows:

   1 TypeTerm     := TypeName
   2               | DQString
   3               | Regex
   4               | QSymbol
   5 QSymbol      := "'" Identifier Spacing

TDL File Interpretation and Conventions

Layout of a type definition

Some parts of a type definition are mandated by TDL syntax, such as the initial identifier, the main operator, and the final dot:

identifier := (definition body) .

The definition body is just a conjunction of terms, maybe with documentation strings, and there is much valid variation in how those terms are arranged. Nevertheless, there are conventional locations for these terms depending on what kind of term they are. For instance, the supertypes are generally listed first, followed by an AVM:

head_only := unary_phrase & headed_phrase &
  [ HD-DTR #head & [ SYNSEM.LOCAL.CONJ cnil ],
    ARGS < #head > ].

If a documentation string is specified, the conventional place is before the AVM:

n_-_ad-pl_le := norm_np_adv_lexent &
"""
<description>N, can modify, locative (place)
<ex>B lives overseas.
<nex>
<todo>
"""
  [ SYNSEM.LOCAL [ CAT.HEAD [ MINORS.MIN place_n_rel,
                              CASE obliq ],
                   CONT.HOOK.INDEX.SORT place ] ].

Or if there is no AVM, before the final dot:

info-str := icons
  """Type for underspecified or "neutral" information structure.""".

Types versus instances

Specifying the text encoding

The text encoding of TDL files can be specified using a special comment on the first line of the file, as is done with many scripting languages. For instance, the following sets the encoding to UTF-8:

   1 ; -*- coding: utf-8 -*-

In some TDL files, attributes specific to the Emacs text editor may be included:

   1 ;;; -*- Mode: tdl; Coding: utf-8; indent-tabs-mode: nil; -*-

Feature interpretation of lists

The < ... > and <! ... !> shorthand for lists ("cons lists") and diff-lists, respectively, correspond to normal attribute-value pairs. The implementation relies on an encoding scheme where the first list item (the list's head) is at the feature FIRST while the rest of the list (the tail) is defined recursively under the feature REST (e.g., REST.REST.FIRST is the third element). The types associated with open and closed lists, and sometimes even the feature names, are configurable by the grammar.

entity

example

LKB config

ACE config

cons-list type

*cons*

(not configurable)

cons-type

open list type

*list*

*list-type*

list-type

closed list type

*null*

*empty-list-type*

null-type

diff-list type

*diff-list*

*diff-list-type*

diff-list-type

list head feature

FIRST

*list-head*

(not configurable)

list tail feature

REST

*list-tail*

(not configurable)

diff-list list feature

LIST

*diff-list-list*

(not configurable)

diff-list last feature

LAST

*diff-list-last*

(not configurable)

For the examples below, I use the values defined in the above table, which are taken from the ERG.

Cons Lists

Regular cons lists may be open (extendable) or closed (fixed-length). The type of an open list as interpreted by, e.g., < ... >, is *list* (rather, the defined open list type), but in hand-written TDL a subtype of *list* is often used, such as *cons*.

   1 ; an empty list is terminated (always empty)
   2 [ ATTR < > ]             =>  [ ATTR *null* ]
   3 ; single item goes on FIRST attribute and REST is terminated
   4 [ ATTR < a > ]           =>  [ ATTR *list* & [ FIRST a,
   5                                                REST *null* ] ]
   6 ; items after the first go on (REST.)+FIRST
   7 [ ATTR < a, b > ]        =>  [ ATTR *list* & [ FIRST a,
   8                                                REST [ FIRST b,
   9                                                       REST *null* ] ] ]
  10 ; an empty list with ... is not terminated
  11 [ ATTR < ... > ]         =>  [ ATTR *list* ]
  12 ; this also works with items on the list
  13 [ ATTR < a, ... > ]      =>  [ ATTR *list* & [ FIRST a,
  14                                                REST *list* ] ]
  15 ; the . delimiter allows a non-*list*, non-*null* value for the last REST
  16 [ ATTR < a . #coref > ]  =>  [ ATTR *list* & [ FIRST a,
  17                                                REST #coref ] ]

Diff Lists

Diff lists are regular lists under a LIST attribute, and LAST points to the last item. Diff lists don't support the unterminated-list functionality of cons lists, but they allow for appending lists of arbitrary size (see GeFaqDiffList).

   1 [ ATTR <! !> ]           =>  [ ATTR *diff-list* & [ LIST #coref,
   2                                                     LAST #coref ] ]
   3 
   4 [ ATTR <! a !> ]         =>  [ ATTR *diff-list* & [ LIST *list* & [ FIRST a,
   5                                                                     REST #coref ],
   6                                                     LAST #coref ] ]

Type documentation

TDL definitions allow documentation strings ("docstrings") before any term in the top-level conjunction or before the terminating dot (.) character:

n_-_c_le := n_intr_lex_entry
"""Intransitive count noun (icn)
<ex>The dog barked.
<nex>Much dog bark.""".

Before docstrings became well-supported, LTDB supported documentation in comments (normally preceding the documented type):

   1 ; <type val="case-p-lex-np-to">
   2 ; <name-ja>承名詞目的格助詞ト
   3 ; <description>case-p-lex-np-woを参照。このtypeは助詞「と」。
   4 ; <ex>部長 と 会う
   5 ; <nex>ゆっくり と 進む
   6 ; <todo>
   7 ; </type>
   8 case-p-lex-np-to := case-p-lex-np &
   9  [SYNSEM.LOCAL.CAT.HEAD.CASE to].

Case sensitivity

Case Sensitive

Case Insensitive

Unknown

Notes: Arguments for case insensitive include shouting (call caps); Arguments for case sensitive include the use of upper case vowels in vowel harmony languages (linguistic representations, not orthography)

Notes for implementation

DocStrings

Multiple docstrings may be present on a single definition, but only the first one encountered on a definition is considered its primary docstring, and implementers are free to store or discard the other doc strings as they see fit. Docstrings on type-addenda should be concatenated with a newline to the previous docstring(s), or appended to a list of docstrings, associated with the type.

Comments

The syntax description above allows for comments anywhere that separating whitespace is allowed (not including those within strings, regular expressions, letter sets, etc.). This includes within a dotted attribute path (e.g., [ SYNSEM #| comment |# . #| comment |# LOCAL ... ]), although grammar developers may want to use this flexibility sparingly.

Open Questions

1. The ^ character is used to signal "expanded-syntax" in the LKB, but is this only used for regular expressions? Are there other expanded syntaxes? Do non-LKB processors support them? (see this thread on the 'developers' mailing list)

2. Are instances distinguishable from types? Are they (other other entities) restricted to having exactly one supertype?

3. Can we use 'status' to identify roots and labels (parsenodes)? Something like

;;
;; parse-tree labels (instances)
;;

:begin :instance :status label.
:include "parse-nodes".
:end :instance.

;;
;; start symbols of the grammar (instances)
;;

:begin :instance. :status root.
:include "roots".
:include "educ/roots-educ".
:end :instance.

Discussions

TdlRfc (last edited 2020-06-05 06:38:36 by FrancisBond)

(The DELPH-IN infrastructure is hosted at the University of Oslo)