Subsections

• The VIT Library Package
• The Verbmobil Interface Term
• Multiple Levels of Information
• Contents of VIT Slots
  • VIT ID
  • Index
  • Conditions
  • Other Slots
• The Prolog Implementation
• Overview

Introduction

The VIT Library Package

The Verbmobil Interface Term (VIT) is one of the still existing ``inventions'' of Verbmobil-1. The same holds for the ADT library package which handles it. No more needs to be said. Well, this report is an update of Verbmobil report 104 [Dorna (1996)] where you may find the first part of the story.

The Verbmobil Interface Term

The VIT is used as a uniform data structure at the interfaces between several software components of the Verbmobil system. It is an encoding of different linguistically motivated pieces of information produced and used in these components.

The contents of a VIT correspond to a segmented utterance in a dialogue turn. This partitioning of turns enables the linguistic components to work incrementally.

The main contents of a VIT are semantic representations. On the other hand, information like morpho-syntax, syntactic tense, semantic sorts, scope, prosody, and the analyzed surface string is also part of a VIT. This information is linked to semantics and can be used for computing semantic tense, for disambiguation of underspecified analyses, for guiding semantic evaluation such as anaphora resolution, for adjacency or linear precedence determination, and for many more.

In a large project like Verbmobil data abstraction is an important basis for the parallel development of different components which should communicate with each other in the end. Hence, also from a software engineering perspective there are a lot arguments for the VIT. Among them there are:

• Because there are abstract access and manipulation operations available the data structures of the VIT can be changed with minor feedback on a software component using it.
• We assume that this is the first time that the result of linguistic components is checked using a kind of analysis structure control. Language-specific on-line dictionaries are used to insure the compatibility between components. A content checker is used to test structural properties and a ``deep'' checker checks even semantically relevant wellformedness. It has been shown that this treatment is well-suited for error detection in components with a rapidly growing linguistic coverage. Furthermore, the complex information produced by linguistic components even make automatic output control necessary.
• The checkers can be used to define a quality rating, e.g. for correctness, interpretability, etc. of the contents of a VIT. Such results are much better to improve a system than common, purely quantitative, measures based on failure or success rates.
• The single VIT representation serves as a common ``language'' (lingua franca ) for discussions. This language can be ``spoken'' by people responsible for technical tasks only as well as others with purely linguistic intentions.

Multiple Levels of Information

The contents of a VIT are filled into the following slots:

Slot Name Description

VIT ID combines a unique tag for a turn segment described by the current VIT and the word lattice path used in its linguistic analysis;

Index a triple consisting of the entry points for traversing the VIT representation;

Conditions labelled conditions describing the possibly underspecified semantic content of an utterance;

Constraints scope and grouping constraints, e.g. used for underspecified quantifier and operator scope representation;

Sorts sortal specifications for instance variables introduced in labelled conditions;

Discourse additional semantic and pragmatic information, e.g. discourse roles for individual instances;

Syntax morpho-syntactic features, e.g. number and gender of individual instances;

Tense and Aspect morpho-syntactic tense combined with aspect and sentence mood information, e.g. used for computing surface tense;

Prosody prosodic information such as accenting and sentence mood.

Beside ``VIT ID'' and ``Index'' all slots are encoded as variable free lists of terms. The lists are used because they are very easy to manipulate. In typical AI languages such as Lisp and Prolog they are built-in and they can be ported easily to other programming languages. In general, the list elements do not introduce any further recursive embedding, i.e. the elements are like fixed records with fields containing constants. The minimally recursive representation was chosen for efficient information access.

The information within the several slots is connected to information in other slots by identical constants which can be seen as bidirectional links. These constants - we call them holes , labels and instances - could be interpreted as skolemized logical variables which denote a node in a graph. I.e., a sharing like this combines information of different nature or source which, therefore, is part of different slots in a VIT.

Contents of VIT Slots

  This section describes the information which can be found in the slots of a VIT. First, we list the terms, and then we explain the possible argument bindings.

VIT ID

  The ``VIT ID'' slot is filled with a VIT identifier of the form 

vitID(SegmentID, WHGString)

SegmentID is a unique marker for each VIT. The Prolog encoding looks like    

sid(TurnNo, Channel, SourceLanguage, BeginTime, EndTime,
Reading, CurrentLanguage, TurnEnd, Sender)

E.g. sid(104,a,ge,7002,7199,1,jp,y,transfer) is an example of a segment identifier with a string representation t104ageb7002e7199r1jpytransfer.

WHGString is a Prolog list which contains edges of a word lattice. This sequence of edges describes the surface string selected by a syntactic/semantic component for producing the current VIT. An edge can have either one of the following forms:   

word(String, WID, ListOfLs)
filler(String, WID)

String is the label of the edge and WID is a unique edge identifier. Both, String and WID, are Prolog atoms. If String is a linguistic information, the word/3 representation is used. For more technical information such as pauses, noice, etc. filler/2 is used.

The list of labels ListOfLs links the edges to some labelled conditions in the Conditions slot. If there is more than one label in a single list, a word was decomposed into several lexemes in semantic analysis. If there are identical labels in different lists, the corresponding words form one labelled condition together. Hence, this encoding reflects an n-to-m mapping between syntax and semantics during linguistic analysis.

Index

  The ``Index'' slot is filled with a term of the form  

index(TopHole,MainLabel,MainInst)

where the hole TopHole, the label MainLabel, and the instance MainInst are the entry points for traversing a VIT.

Conditions

  The contents of the ``Conditions'' slot are language specific. In general, it contains terms of the form

Functor(Label, Arg2, ..., Argx)

called labelled conditions. The semantic entities are, e.g., predicates, roles, operators, and quantifiers. The first argument - sometimes called base label - is always a unique identifier for such an object. The semantic variables for labels and markers, such as events, states and individuals, are skolemized with special constant symbols, e.g. l1 for a label and i1 for a state.

The labeling of semantic conditions is very useful since the recursive embedding of argument structure, operator scope, etc. is no longer syntactically represented in a recursive data structure but achieved through the interpretation of additional labeling constraints. In this respect, label arguments act as pointers to the corresponding arguments. Additionally, all these special constants can be seen as pointers for adding or linking information within and between multiple slots of the VIT.

The set of all possible conditions for one language is defined by its lexicon. The ADT package supports on-line lexicons for arbitrary languages. See Appendix  for the expected format and for hooks to extend a language specific lexicon. Examples of such lexicons can be found in vitSemLex.pl for English, German, and Japanese.

.

Other Slots

  The information located in the rest of the slots is given in the following table:

Info Description Slot Name

ana_ante(Inst,ListOfIs)  anaphoric antecedents Discourse

c_class(Label,CClass)  coordination type Discourse

cas(Inst, Case)  syntactic case Syntax

dialog_act(Label, DialogAct)  dialogue act information Discourse

dialog_phase(DialogPhase)  dialogue phase Discourse

disc_honor(Inst,Inst)  honorifics (for Japanese) Discourse

disc_role(Inst,DiscRole)  discourse role Discourse

disc_stat(Label,Status)  discourse status Discourse

discourse_function(L,DiscFunc)  pragmatic discourse relation Sorts

dir(Label, YesNo)  (non)directional preposition Discourse

eq(Label,LabelOrHole)  equality constraint Constraints

gend(Inst, Gender)  morpho-syntactic gender Syntax

in_g(Label, GroupLabel)  label is in group Constraints

leq(Label,Hole)  scope/subordination constraint Constraints

nom_ante(Inst,ListOfIs)  anaphoric reference Discourse

num(Inst, Number)  morpho-syntactic number Syntax

pcase(Label,Inst,Atom)  subcategorized preposition Syntax

pers(Inst, Person)  person Syntax

prontype(I, PRef, PType)  type of a pronoun Discourse

pros_accent(Label)  prosodic accent Prosody

pros_boundary(Label)  prosodic (b3) marker Prosody

pros_emph(Label)  emphasis marking Prosody

pros_mood(Label, PMood)  prosodic mood Prosody

pros_probab(Label, Prob)  prosodic accent probability Prosody

ref_ante(Inst, Inst)  anaphora resolution Discourse

rel_ante(Inst,ListOfIs)  anaphoric reference Discourse

s_class(Label, SClass)  disambiguation class Sorts

s_sort(Inst, Sort)  sortal restriction Sorts

sem_focus(Label, LabelOrHole)  focus resolution Constraints

subj_honor(Inst,Inst,Honor)  subject honorific marker (Japanese) Syntax

syn_cat(Label,Atom)  syntactic category Syntax

ta_aspect(Inst, Aspect)  aspectual information Tense and Aspect

ta_mood(Inst, TMood)  mood Tense and Aspect

ta_perf(Inst, Perf)  perfect Tense and Aspect

ta_tense(Inst, Tense)  surface tense Tense and Aspect

topic(Label,GroupLabel,Topic)  topic information Discourse

unbound(LabelOrInstOrHole)  unbound markers Discourse

The argument values of the terms above and of the predicates described in the next sections are given in the following table:

Argument Description or list of values

Argx Hole, Inst, Label, Atom, or a Prolog list of the named Args

ArgType np, acomp, comp, xcomp

Aspect progr, nonprogr plus prospective, terminative for Japanese only

Atom Prolog atomic

BeginTime unsigned integer

Case ;(Case,Case), nom, gen, dat, acc, noc(ase)

Check ground, shape

Channel a, b, c, d, e, f, g, s, t

Class ambig, disc, grad, mod, mood, noun, quant, verb

CurrentLanguage see Language

CClass m, n, q, s, s

DialogAct ;(DialogAct,DialogAct), ~(DialogAct), accept, backchannel, bye, clarify, close, commit, confirm, control_dialogue, defer, deliberate, deviate_scenario, dialogue_act, digress, exclude, explained_reject, feedback, feedback_negative, feedback_positive, give_reason, greet, inform, init, introduce, manage_task, offer, politeness_formula, promote_task, refer_to_setting, reject, request, request_clarify, request_comment, request_commit, request_suggest, signal_non_understanding, suggest, thank

DialogPhase hello, opening, negotation, closing, good_bye

DiscFunc attitude, check, coherence, edit, emphasize, exemplify, hesitate, indifferent,, known, negative, nonpragmatic, pop, positive, prefer, push, repair, revised, smoothen, structure, surprise, uptake

DiscRole he, sp

EndTime unsigned integer

Functor Prolog functor

Gender ;(Gender,Gender), fem, masc, neut

GramFunc subj, obj, obj2, nof(unction) Hole or H

Honor minus, plus

Info any valid term encoding information in a VIT

Inst or I Prolog atom of the form i[a-z][0-9]*

Label or L Prolog atom of the form l[a-z][0-9]*

Language de, en, ge, jp

ListOfLs Prolog list of Label elements

ListOfIs Prolog list of Inst elements

TMood ind, conj, imp

Number ;(Number,Number), sg, pl

Perf perf, nonperf

Person ;(Person,Person), 1, 2, 3

PRef sp, he, sp_he, third, third_he, top

$$0 \leq i \leq 100$$ Prob

PType refl, std, refl_std, recip, imp, event, event_std, demon, demon_event, zero (and intersent for Japanese)

PMood decl, prog, quest

Reading integer

SemRole arg1, arg2, arg3, refl

Sender Prolog atom which is the name of a Verbmobil component

String Prolog atom encoding a L^ATEX string

SClass sa, fp, mp, dra

SMood decl, imp, quest, decl_quest, decl_imp, imp_quest, decl_imp_quest

SlotName 'Conditions', 'Constraints', 'Tense and Aspect', 'Index', 'Prosody', 'Discourse', 'Sorts', 'VIT ID', 'Syntax'

Sort ;(Sort,Sort) or &(Sort,Sort) or ~(Sort) or abstract, action_sit, agentive, animal, anything, communicat_sit, entity, field, food, geo_location, human, info_bearer, info_content, institution, instrument, location, meeting_sit, mental_sit, move_sit, nongeo_location, object, position_sit, property, situation, space_time, substance, symbol, temp_sit, temporal, thing, time, vehicle

SourceLanguage see Language

Status old, new, inferred_old, contrast_new

Stream a valid Prolog stream object

Tense pres, past, future

Topic fr, mo, vf, wa

TurnEnd n, y

TurnNr unsigned integer

VIT Verbmobil Interface Term (not necessarily ground)

YesNo yes, no, yesno

The Prolog Implementation

  In Prolog, the VIT is implemented as a term of arity 9 named vit. The name and arity of this realization should be of no interest if the access to a VIT is always handled by the ADT package (or a similar abstraction). If this is the case, we are not restricted to this implementation in future. As already pointed out, in general, the information is encoded in lists and the elements are terms. This data structure is very flexible in adding and removing information, i.e. the terms.

The output of each component dealing with VITs has to make sure the following properties:

• A VIT is (logical) variable free, i.e. all variables are skolemized.
• A VIT contains only valid information, i.e. in general, lists of terms with atomic arguments.
• The syntax of a term in a VIT is unique with respect to the slot it belongs to. I.e. the functor and arity are sufficient to uniquely determine the adequate slot. Additionally, for labelled conditions the functor and arity are sufficient to uniquely determine the semantic class it belongs to.
• There is only a fixed number of possible terms allowed in a VIT which are restricted to a fixed number of possible argument values (enumerations) or to specific types (e.g., integers).

For describing the predicates of the ADT package we use the ``standard'' notation for call patterns (aka mode information):

+  

the argument is expected to be instantiated (not necessarily ground) and will not be changed during processing of the called predicate;

-  

the argument is expected to be a variable and will be bound during processing of the called predicate;

?  

the argument can be instantiated when calling the predicate and/or will be bound during processing of the called predicate.

The ADT package is realized as a Prolog module named vitADT which exports the predicates described in the following sections. For further remarks on the usage see Appendix .

Overview

The rest of this documentation is organized as follows. In Section  we present the predicates for constructing a new VIT and filling it with information. In Section  and Section  we outline predicates for information access and those for deleting information, respectively. Section  informs about predicates for checking the contents of a VIT. In Section  we show predicates for printing a VIT. Miscellaneous predicates are described in Section .

Appendix  explains how to get and install the ADT package and Appendix  shows how to use it. Appendix  introduces the term conversion package atom2term which comes with the library package distribution. Finally, Appendix  sketches briefly the contents and possible extensions of the on-line dictionaries contained in vitSemLex.pl and vitValues.pl.

For each predicate presented in this document we give the call pattern(s) and a brief description sometimes including an example.