AN APPROACH TO MULTILEVEL SEMANTICS 
FOR APPLIED SYSTEMS 
Alberto Lavelli Bernardo Magnini Carlo Strapparava 
IRST, Istituto per la Ricerca Scientifica e Tecnologica 
I - 38050 Povo TN, Italy 
e-mail: magnini@irst.it 
Abstract 
Multilevel semantics has been proposed as a 
powerful architecture for semantic analysis. We 
propose a methodology that, while maintaining the 
generality of the multilevel approach, is able to 
establish formal constraints over the possible ways 
to organize the level hierarchy. More precisely, we 
propose a "strong" version of the multilevel 
approach in which a level can be defined if and only 
if it is possible to characterize a "meaningfulness" 
notion peculiar to that level. Within such an 
architecture each level reached during the analysis 
computes its meaningfulness value; this result is 
then handled according to modalities that are peculiar 
to that level. 
The component described in this paper was 
designed to be portable with respect to the 
application domain and so far has been tested as the 
semantic analysis component of two multimedial 
dialog systems, ALFresco and MAIA. 
1. Introduction 
Multilevel semantics has been proposed \[Scha, 1983\] as 
a powerful architecture for semantic analysis. In this 
approach, interpreting a natural language sentence is a 
multi-stage process, which starts out with a high-level 
meaning representation that reflects the semantic structure of 
the sentence rather directly. Then translation rules, which 
specify how the language-oriented semantic primitives relate 
to those used at deeper levels of analysis, are applied. One of 
the advantages of the multilevel approach is that it allows a 
natural decomposition of complex tasks and the functional 
modularization of semantic analysis. However, when 
multilevel architecture is used in concrete applications, a 
simple functional approach does not solve the problem of a 
clear definition of the semantics for each level. This fact is 
evident for applied systems whose semantic component 
must deal with many linguistic phenomena (e.g. lexical and 
structural ambiguities, quantifier scoping, anaphorical 
references, discourse topic and focus, referent retrieval, etc.). 
In such systems the definition of the semantics for a level 
has at least two advantages: (i) modules for specific 
phenomena could be easily introduced within the appropriate 
level, provided that the module functions contribute to the 
definition of the semantics for that level; (ii) a better 
understanding of the semantic analysis would be allowed: 
particularly, when a sentence is rejected at a certain level, it 
would mean that the semantic constraints for that level have 
been violated. 
In this paper we suggest a methodology that, while 
maintaining the generality of the multilevel approach, is 
able to establish formal constraints over the possible ways 
to organize the level hierarchy. More precisely, we propose 
a "strong" version of the multilevel approach in which a 
level can be defined if and only if it is possible to 
characterize a "meaningfulness" notion peculiar to that level. 
Within such an architecture each level reached during the 
analysis computes its meaningfulness value; this result is 
then handled according to modalities that are peculiar to that 
level. 
We shall show how our approach to multilevel 
semantics is concretely applied to organize the semantic 
component developed by the NLP group at IRST; this 
component is currently responsible for semantic analysis in 
two dialog systems, ALFresco and MAIA. At present two 
levels are included in the semantic component and they will 
be described in detail: the lexical level and the logical- 
interpretative level. At the lexical level the meaningfulness 
is defined by the consistency notion, which is computed by 
means of the lexical discrimination module; this module 
tries to select only the sentence readings meaningful in a 
given Domain Model (DM). When the propositional content 
of the sentence is proven to be consistent, the semantic 
representation produced by this level is passed to the next 
one; otherwise, if consistency cannot be proved, the whole 
sentence is rejected. At the logical-interpretative level the 
meaningfulness is defined by means of the validity notion, 
which is satisfied when referents for the sentence are 
identified. Three modules interact at this level: the 
quantification module, which finds the correct interpretation 
of the quantifiers, resolving possible scoping ambiguities; 
the topic module, which organizes the mentioned referents; 
the interpretation module, which identifies the part of the 
sentence to extensionalize and is responsible for referent 
retrieval. At this level, when validity cannot be proved, a 
special pragmatic procedure is activated. 
Section 2 surveys a few relevant approaches to 
multilevel semantic analysis. In Section 3 the formal 
requirements for the "strong" multilevel semantics version 
are introduced. The architecture and the functional modules 
of the two levels of the semantic component we have 
developed are described in Sections 4 and 5. Finally, Section 
6 deals with tile two systems in which the semantic 
1 7 
  
 17 
component has been used and Section 7 outlines some 
future developments. 
2. Multilevel Semanlics Applied 
One of the first and most direct multilevel-based systems 
is the BBN spoken language system \[Boisen et al., 1989\]. 
At every level of analysis, the meaning of an input utterance 
is represented as an expression of a logical language; the 
languages used at the various levels of analysis differ in that 
at every level the descriptive constants are chosen so as to 
correspond to the semantic primitives assumed at that level. 
At the highest semantic level, the meaning of an input 
utterance is represented as an expression of the English- 
oriented Formal Language (EFL). The constants of EFL 
correspond to the descriptive terms of English. An 
important feature of EFL is that descriptive constants are 
allowed to be ambiguous. The logical language used at the 
domain-dependent level of representation is called the World 
Model Language (WML). This is an unambiguous 
language, with an ordinary model-theoretic interpretation. Its 
constants are chosen to correspond to the concepts that 
constitute the domain of discourse. During the crossing of 
the EFL and the WML level (when domain dependent 
rewriting rules are called), the discrimination process is 
carried out. A type checking mechanism provides acceptance 
only for interpretations for which a domain knowledge 
compatible type has been computed. A further step of 
translation occurs when the WML is translated into DBL 
(DataBase Language) used to access the database to retrieve 
appropriate answers. 
While having sound theoretical foundations, the main 
drawback of this approach is that it postpones semantic 
discrimination until domain knowledge is available; in the 
meantime, a complete sentence representation is built for 
each analysis the parser produces. However, IRUS-II \[Ayuso 
et al., 1989\], an applicative system also developed at BBN, 
confirms that in a real system it is useful to connect the 
discrimination process to the parser. It implements a rule 
system that translates each syntactic constituent directly into 
a WML form, skipping the domain independent level of 
representation. While this solution improves system 
efficiency, lexical discrimination is carried out by domain 
dependent rules in a way that limits system modularity. 
Another system with a clear distinction between the 
domain independent and the domain dependent level is 
XTRA \[Allgayer et al., 19891. However, in this case at each 
level .the same language (i.e. the knowledge representation 
language SB-ONE) is used. The domain independent level, 
called Functional-Semantic Structure (FSS), is intended as 
an intermediate structure that incorporates linguistic 
knowledge, substantially invariant in respect to the 
particular application domain. On the contrary, the domain 
dependent level, called Conceptual Knowledge Base (CKB), 
is necessary to adequately model the relations of the 
underlying expert system. In XTRA it is necessary that each 
analysis produced by the parser is consistent with the FSS 
level: this is achieved by means of a classification of the 
sentence instance with the SB-ONE mechanisms (the 
realizer and the matcher). If the classification succeeds, the 
analysis goes on to the CKB level, otherwise the syntactic 
analysis is rejected. In this approach the discriminatiol 
process is profitably anticipated, and a powerful (eve1 
though computationally expensive) consistency checkinl 
mechanism is provided. 
Both systems exploit the difference between knowledg 
about the application domain and knowledge that i 
independent of the particular domain (e.g., linguisti 
knowledge). Although this distinction is relevant fc 
allowing portability to different application domains, th 
semantic component described here focuses on the effecl 
that domain dependent knowledge has on the type checkin 
mechanism. 
To make the problem clearer, let us consider ho~ 
domain knowledg e is exploited in the systems ju~, 
described. In the BBN spoken language system the typ 
checking is carried on by means of domain knowledge; o 
the other hand, within the XTRA system the discriminatio 
process is based only on domain independent knowledge. W 
think that an effective discrimination process should also b 
based on the application domain, it being unclear how t 
assign a proper meaning to a sentence without having fixe 
a particular context. Moreover, it seems useful to considc 
lexical discrimination as an incremental process: i 
discrimination works in parallel with the parser, it i 
possible to discriminate over single syntactic phraset 
checking the semantic content of each phrase. 
From the previous remarks, it can be noted that system 
that employ the multilevel semantics approach can assig 
the same functionalities to different levels. Hence, it coul 
be useful trying to define the relations among each level in 
"stronger" way, facing the problem of coherenc 
maintenance. 
3. Definitions of Meaningfulness 
We have seen that in a multilevel semantics approac 
the main idea is to divide different functionalities int 
distinct levels. We propose a "strong" approach to such 
methodology in which for each level the definition of 
semantics is required. This is achieved by means of th 
assignment of a proper meaningfulness notion that defin~ 
the semantic behavior of the level. In other words a level i 
a multilevel semantics hierarchy can be identified, if an 
only if it is possible to characterize a meaningfulness notic 
peculiar to that level. We have defined theoretically such 
notion for two levels: the lexical level and logica 
interpretative level (called consistency and validit 
respectively). 
Let T be a theory of types that models our domain, l 
our multilevel semantics the notion of consistency is meal 
to demonstrate that an expression, representing tl~ 
propositional content of a sentence, has type; i.e. given 
expression w, it means to assign a type, if possible, to 
according to our type system. An expression has n 
meaning at the lexical level, if the type checking fails. 
Validity, i.e. the meaningfulness at interpretation leve 
means to give a description of the objects of the tyl: 
suggested by the lexical level. Such a description can be i 
terms of relations, sets or intensional expressior 
(mandatory for infinite denotations). An expression has r 
meaning at the logical-interpretative level if such 
  
 18 
description cannot be found. 
As the meaning of a sentence is always relative to a 
level in the multilevel architecture, every level manages the 
acceptance or the rejection of a sentence in a different 
manner. As examples: 
(1) A mule paints a fresco 
The components of the sentence have the following 
types: 
a mule : Mule, a fresco : Fresco, 
to paint : Painter --9 Painting. 
Given the fact that "mules cannot paint" (only painters 
can), the type checking mechanism fails to assign an 
appropriate type and this causes the meaningfulness for the 
lexical level not to be satisfied. 
(2) Show me a work painted by all the painters born in 
Florence 
Sentence 2 satisfies the lexical level, but not the logical- 
interpretative one, because no description of the referents of 
the sentence can be proposed, i.e. there is no painting 
painted by all the painters born in Florence. 
Once the functionalities of the levels are theoretically 
stated, the implementative choices can be very different and 
subject to criteria of portability. Type checking can be made 
using logical formalisms such as typed ~,-calculi or 
intensional logics (possibly exploiting Curry-Howard's 
isomorphism between typed ~.-terms and intuitionistic logic 
\[Hindley and Seldin, 1986\]). The interpretation level can 
retrieve the referred elements using functional applications 
or some algebraic formalisms. However, these approaches, 
although well founded, may not be the right ones from an 
implementative point of view, especially for large integrated 
systems. For example one has to define 'a priori' a theory of 
admissible types but when the domain changes, the theory 
does too. Another way is to use a hybrid knowledge 
representation system. As will be clear in the next section, 
we refer to a terminological component (Tbox) in order to 
obtain the type checking and to an assertional component 
(Abox) in order to retrieve the relations that verify the 
analyzed expression. This choice allows us to parameterize 
the type checking according to the knowledge representation. 
Indeed the portability of the modules encourages this 
alternative. Another possibility (to be explored) is to use a 
data base instead of the Abox, exploiting relational data 
theories. 
4. Lexical Level 
The semantic component (see Figure 1) interacts with 
both a parser and a hybrid knowledge representation system 
that includes the domain knowledge. As we have already 
mentioned, the semantic component consists of two levels 
and each level includes one or more specialized modules. In 
the following we will give a description of the 
functionalities of the various levels and modules of the 
semantic component. 
The lexical level \[Lavelli and Magnini, 1991\] 
incrementally interacts with the parser: whenever the parser 
tries to build a (partially recognized) constituent, the 
discrimination module is triggered to check the consistency 
of the semantic part of such a constituent. 
Input Sentence 
Lexical Level ~ "(Parser) 
Discrimination consistency 
module 
iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii!iiiii!iii!' i r -, 
iiii Logical-Interpretative Tbox 
::~::~ Level 
!i!i ' (Quantification) Abox 
module ~'~ J 
Q Interpretation m o,e )-- v i ty 
Oomain 
.... Model Topic 
module 3 
Figure 1: Sketch of the semantic process 
4.1. Lexicon 
The discrimination module uses semantic information 
from two different sources: lexical entries (which are domain 
dependent) and phrase-structure rules (which are domain 
independent). The representation produced by this module 
constitutes the input for the quantification module (at the 
logical-interpretative level) and is still neutral with respect 
to quantifier scopings. 
Each lexical entry, along with the usual syntactic 
information (such as the lexical category of the word, the 
specification of the subcategorization frame of the entry, the 
superficial linguistic function that each subcategorized 
element holds) specifies a semantic representation and a 
mapping between syntactic functions and semantic 
functions. In such a way, within the semantic representation 
the syntactic distinction between the word complements (i.e. 
the arguments) and its adjuncts (i.e. its modifiers) is 
preserved. 
As an example consider the simplified lexical entry for 
the verbal form "dipinse", painted (past tense) (see Figure 
2). Morphological analysis enriches the information 
associated with the root and is able (for example in the case 
of passive) to change the mapping between linguistic 
functions and semantic functions. The semantic part of the 
lexical entry is built using the domain model knowledge 
(see Section 7 for a discussion on the portability problem) 
and it includes one (or more) semantic descriptions (this 
allows words with the same syntactic behavior, but different 
semantics, to be dealt with). Each semantic description 
contains the name of the DM concept (paint) associated 
  
 19 
with the word, along with its roles, which have a syntactic 
realization as arguments of the word and their restrictions 
(in this case, agent with restriclion painter and goal with 
restriction painting). 
dipinse 
category: V 
lingfunctions: ((subj agent) (obj goal)) 
<other syntactic information> 
semantics: ((paint ((agent painter x) 
(goal painting y))) lex 
Figure 2: Lexical entry for "dipinse": painted. 
As for the rules, they also include both a syntactic and a 
semantic part. In the semantic part, the consistency is 
computed and the construction of the semantic 
representation is carried out. During this process, possible 
ambiguities taken from lexical items are reduced. 
:ion 
i i I t il t 
I I < Cappella Scrovegni > / I I < paint#1 > t j t 
• . I i t spaual-locanon t t 
< Giotto > '~ II / < Trento > 
Abox < Cacciata > < vaaova > 
Figure 3: A fragment of the DM used in ALFresco. 
4.2. Consistency checking 
We define the consistency check operation such that it 
succeeds if selectional-restriction (i.e. the concept that 
represents the selectional restriction of a given argument 
position) denotes a concept that is compatible with the 
concept that semantic-head (i.e. the concept associated with 
the constituent which has to fill such a position) denotes. 
There exist several possibilities to check the 
compatibility between two concepts within a terminological 
hierarchy. Within the JANUS system \[Weischedel, 1989\] the 
consistency is implemented by means of a double 
subsumption check that guarantees success both when 
semantic-head is a descendant of selectional-restriction and 
when it is an ancestor. This double subsumption test does 
not consider the cases, sometimes relevant, in which 
semantic-head is a brother concept of selectional-restriction 
(e.g. "Has a sculptor painted a fresco?"); this case 
recursively extends to all the cases in which semantic-head 
is a brother either of a descendant or of an ancestor for a 
selectional-restriction (e.g. "Which object did Giotto 
paint?"). This case is slightly more complex than the 
others. In fact, while it is always true that along the ISA 
hierarchy there can be a non-empty intersection between two 
concepts, this is not true for concepts that are brothers. If an 
explicit disjoinmess is placed between two brother concepts, 
there cannot be a common intersection and the consistency 
procedure must fail; otherwise it is assumed that a common 
intersection can exist, and the consistency-test procedure 
will succeed. KR languages with disjointness are usually 
provided with a specific predicate holding between two 
concepts when their intersection is empty. It is worth 
noting that this predicate includes all the subsumption cases 
among concepts, in which cases it is always false. 
Now we will illustrate how the whole process works 
using Sentence (3) (in the rest of the paper, all the examples 
refer to the DM knowledge in Figure 3; we will use 
"concept" characters to indicate DM objects): 
(3) 'Mostrami tutti gli affreschi dipinti da Giotto in un 
monumento di Padova' 
Show me all the frescoes painted by Giotto in a 
monument of Padova 
In this sentence there is a typical case of ambiguity, that 
of the preposition 'di' (of); at least two senses for 'di' are 
possible in DM: the spatial interpretation, in which the 
mapping is to the spatial-location role, and the 
temporal interpretation, in which the mapping is to the 
temporal-location role. The selection of the right 
interpretation (the spatial one) is carried out through the 
application of the consistency check between the argument 
selectional-restrictions (the domain and the range of a role) 
and the semantic-head that tries to fill the position. In this 
case the temporal interpretation is rejected (it does not 
satisfy the meaningfulness notion for the lexical level) 
because the range restriction (time-period)is not 
consistent with the proposed semantic-head (Padova). 
4.3. First logical form 
The final result of the lexical level is a form that uses a 
predicate-argument notation that allows abstracting from 
time and context. Omitting for the moment the intensional 
aspects, four relevant constructs for the resolution of 
quantifiers and definite/deictic referents are: 
(complex-term <features> 
<quan tifier><variable><restriction>) 
  
 20 
(ref-term <features><variable>< restriction>) 
(demonst r-term<features><variable><restriction>) 
(pronoun-term <features> 
<variable><pred-restriction>) 
A complex-term represents a quantified NP (see Figure 
4). A ref-term represents a definite NP. It plays an important 
role at interpretation level (see Section 5.3). A demonstr- 
term has the task of representing a demonstrative NP. The 
representation has to take into account the possible 
multimediality that the system treats at this level (the touch 
on the touchscreen for a deictic reference). A pronoun-term 
represents a pronoun. The lexical level gives a suggestion 
with <pred-restriction> on the type of semantic restriction 
that the bound variable can have. Then this information will 
be used by the interpretation module. The <features> keep 
syntactic information of the NP ready for use in the 
interpretation module. 
(show hearer 
(complex-term all x 
(and (fresco x) 
(paint Giotto x) 
(spatial-location x 
(complex-term indef y 
(and (monument y) 
(spatial-location y Padova))))) 
speaker) 
Figure 4: Output of the lexical level for Sentence (3). 
The resulting form produced by the lexical level for 
Sentence (3), omitting the <features> information, is shown 
in Figure 4. 
5. Logical-Interpretative Level 
At this level validity of the sentence is checked using the 
knowledge in the DM Abox. Verifying the validity of a 
logical form and producing the correct interpretation is not a 
trivial task. We want the semantic interpreter to be 
independent of the domain of the knowledge representation 
system and of the different media through which a linguistic 
expression can be built. This process involves the 
interaction of the three modules of this level shown in 
Figure 1 \[Strapparava, 1991\]. 
5.1. Quantification Module 
Within the quantification module, an algorithm for the 
resolution of quantifier scopings generates all possible 
readings and for each quantifier it shows its range over the 
rest of the sentence. However, to get an acceptable number 
of readings (possibly only one), the scoping generation 
algorithm, which takes advantage of the idea of Cooper 
storage, needs some heuristics based on linguistic/semantic 
knowledge. These rules must be seen as a whole, i.e. they 
strictly interact with each other. Moreover they suggest a 
disambiguation, they do not always ensure it. Some rules 
can be: 1) lexical relevance of the quantifiers; 2) syntactic 
position of quantified NPs; 3) scope markers; 4) 
distributive/collective semantics of predicates. The readings 
are put in order of soundness according to a hierarchy of 
rules. 
The scoping resolution algorithm produces a second 
logical form in which all complex-terms are resolved, 
making their scope and body explicit. In this logical form 
for each quantifier a quantified formula appears with the 
following structure: 
(quant var restriction body) 
For example, the reading for Sentence (3) in which for 
each fresco there exists a monument that includes it, is 
shown in Figure 5. 
(all x (indef y (and (monument y) 
(spatial-location y Padova)) 
(and (fresco x) 
(paint Giotto x) 
(spatial-location x y))) 
(show hearer x speaker)) 
Figure 5: Second logical form for Sentence (3). 
5.2. Interpretation Module 
The interpretation of the logical form built by scoping 
resolution makes up a level in which the validity of a 
sentence is detected and eventually the relative referents are 
retrieved (possibly interacting with the topic module in order 
to get referents for linguistic expressions such as definite 
NPs and personal pronouns). The expressions are mapped 
into the KR assertional language. The main task of the 
interpretation module is the interpretation of the logical 
form operators, giving a set of possible candidates that 
logically satisfies the sentence for each NP. The results are 
then notified to other modules of the system (i.e. the 
pragmatic component). 
The interpretation of the operators includes the quantifier 
interpretation (existential, universal, numerals and natural 
quantifiers). The restriction of a quantified formula is 
calculated and the result is logically verified in the body 
according to the semantics of the quantifier operators. Since 
there may be an arbitrary nesting of quantifiers in the second 
logical form of a sentence, the algorithm has to provide an 
arbitrary deep recursion of such functionalities. (Indeed the 
interpretation module has other important tasks. One of the 
improvements under development consists of embedding 
intensional aspects into the logical form. These intensional 
aspects tend to extend the characteristic of an extensional 
logical form by allowing references to time and contexts 
(indexicals) \[Forbes, 1989, Stallard, 1987\]. They would also 
include the possibility of interpreting certain NPs along the 
attributive/referential dimension). 
For a detailed description of the algorithms of the 
logical-interpretative modules see \[Strapparava, 1991\]. Now 
we want to focus on the interpretation of quantifier 
operators. According to the semantics of these operators the 
interpretation module checks the validity of a sentence. 
  
 21 
5.2.1. Semantics of quantification operators 
The notation that will be used in discussing the 
semantics of quantification operators is given below: 
pred\[x\] indicates a well-formed form in which the 
variable x appears free; 
- ext(kx.p\[x\]) indicates the extension of p in a 
representation domain DM; 
~(tx.p\[x\]) indicates the set of the parts of the 
extension denoted by p; 
- be I a set, ITI indicates its cardinality. 
We shall show how semantics is assigned to the 
quantification operators in the logical form. 
As seen above a quantifier is syntactically represented 
with the wff 
(quant x rest rict ion\[x\] body\[x\]) 
that has a semantic interpretation 
{ a e \]~quant(kx.rest rYctYon\[x\]) 
I body\[x/a\] is verified in DM} 
where ~quant(kx.restrict ion\[x\]) is appropriately 
defined for each treated quantifier. 
The quantification operators that can appear in the 
logical form are universal and existential quantifiers, wh- 
operators, natural quantifiers such as numerals (two, 
three...), exception operators (all except three ...), vague 
operators (many, several, most of...). As an example we 
show how semantics is assigned lo the quantifier 'many'. 
About 'moiti' (many) there can be two attitudes: either 
one excludes this type of quantification by an extensional 
treatment \[Keenan and Stavi, 1986\] or one tries to get what 
'many' means in a fixed context \[Barwise and Cooper, 
1981\]. In our approach this second consideration was 
followed. Therefore 
~many(kX.p\[x\]) = 
{ P \]P(Z,x.p\[x\]): Ipl = kl\[ext(~.x.p\[x\])\]l } 
where the multiplier k may be fixed, for example 0.6, or 
may depend on pragmatic aspects or on inferences on the 
semantic structure of the dialog. 
Also to interpret other vague operators extensionally, it 
is necessary to make a stipulation of cardinality (fixed or 
dynamic), depending on the inferential capabilities of dialog 
structure at our disposal. For example almost all, most of 
may be interpreted as semantically similar to 'except at 
most K', where for K considerations similar to those made 
for 'many' hold. 
For example, we can apply the operators to sentence (3). 
According to the domain DM, if the interpretation module 
may construct a and a', i.e. to construct a description of the 
referents of the sentence, the validity for sentence (3) holds. 
If the sentence were 'Mostrami tutti gli affreschi dipinti da 
Giotto in un monumento di Trento' (Show me all the 
frescoes painted by Giotto in a monument of Trento) it 
would be consistent, but the logical-interpretative level 
would have found it not valid (because there are no frescoes 
by Giotto in Trento). The interpretation module would not 
be able to construct a description of the referents of the 
sentence. 
5.3. Topic Module 
The ref-terms, demonstr-terms and pronoun-terms are 
treated specially. The demonstr-terms coming from a deictic 
gestuality (i.e., in our systems a touch on a touchscreen; see 
Section 6) contain the entities to which the user intended to 
refer. These are passed to the interpretation module to verify 
the semantic consistency. The demonstr-terms without 
touch, the pronoun-terms and some ref-terms are resolved 
with strict interaction with the topic module. The topic 
module organizes the mentioned referents so that it offers 
plausible candidates for these operators and the interpretation 
module verifies their semantic soundness. For a detailed 
description of the topic module, see \[Samek and 
Strapparava, 1990\]. First of all, the constants in the logical 
form (in our example: Giotto and Padova) are passed to the 
topic module. Later on the topic module is asked to give a 
set of probable candidates for the terms in the logical form 
coming from a deictic gestuality and from the terms coming 
from pronouns. The interpretation then will test their 
validity. 
{ a' s ~exist 
where a is 
{ az \]Pall( 
\, 
I ky.(and (monument \[y\]) 
(spatial-location \[y\] Padova) l DM verifies 
(and (fresco a) 
(paint Giotto a) 
(spatial-location a a' )) 
} 
kx.(indef y (and (monument y) 
(spatial-location y Padova)) 
(and (fresco \[x\]) 
(paint Giotto \[x\]) 
(spatial-location \[x\] y))) 
) \]DM verifies (show hearer a speaker)) } 
Figure 6: Quantification operators applied to Sentence (3). 
  
 22 
6. Application on Different Complex 
Systems 
The semantic component described in this paper has been 
used within two different prototypical dialog systems (i.e 
ALFresco and MAIA). 
ALFresco is an interactive system for a user interested in 
frescoes. It is connected to a videodisc unit and a 
touchscreen. The videodisc includes images of Fourteenth 
Century Italian frescoes and relevant monuments and 
hypertext includes art critics' comment. A general 
description of the functionalities and finalities of the 
ALFresco system can be found in \[Stock, 1991\]. 
MAIA is the IRST global project. It is conceived as an 
integration of components being developed here in different 
fields of AI (speech recognition, natural language, KR, 
vision, reasoning, etc). It consists of both a mobile part (a 
robot moving in the corridors of the institute) and a central 
part (a kind of "concierge" with whom a visitor may enter 
into a dialog about the institute). The tasks of the concierge 
are: (i) giving information about researchers' activities and 
institute organization; (ii) supervising the robot's activities; 
(iii) interacting with an electronic librarian in order to find 
relevant books. The initial paradigm for the concierge 
interaction is related to that of ALFresco, but of course the 
situation and media are different. As the project evolves 
natural language dialogs will also include direct interaction 
with the robot (whose role is to accompany the visitor to 
some office or deliver parcels) and an integration with 
speech recognition and synthesis. Within this more complex 
situation, the NLP component has to increase its 
capabilities in order to cope with aspects such as multiple 
access to information and interaction with the robot planner. 
Both systems have a common architecture design and 
have been implemented in CommonLisp within the Medley 
environment running on Sun 4. The main components 
interacting with the semantic component described here are a 
parser and a hybrid knowledge representation system. Both 
for ALFresco and MAIA the parser WEDNESDAY 2 is used 
\[Stock, 1989\], a chart-based parser for the Italian language 
that can cope with complex sentences, idiomatic 
expressions, ellipsis, and so on. 
As for knowledge representalion, in ALFresco the YAK 
system \[Franconi, 1990\] is used, while in MAIA the 
LOOM system \[McGregor and Bates, 1987\] is used. 
7. Conclusions and Future Work 
We have presented an approach to multilevel semantics 
that was exploited in the development of two semantic 
levels for a dialog system architecture: the lexical level and 
the logical-interpretative level. The suggested methodology 
is able to establish formal constraints over the hierarchy by 
means of a local meaningfulness notion. Such a notion was 
defined for the lexical and logical-interpretative level, 
specified as consistency and valiclity respectively. Then how 
the functionalities of each level realize their own semantic 
definitions was explained in full detail. Finally two 
systems, ALFresco and MAIA, that use the semantic 
component were described. 
Future developments of our work concern the issue of 
portability to different application domains. While the 
general inference mechanisms employed by both the lexical 
and the logical-interpretative level are designed to be 
domain-independent, the semantic lexicon contains 
information strictly connected with the domain of 
interaction. To (at least partially) automatize the 
construction of this semantic lexicon (given a particular 
DM), the possibility of using an approach similar to the 
Upper Model used by the PENMAN text generation system 
\[Bateman et al., 1990\] is being investigated. The Upper 
Model establishes a level of linguistically motivated 
knowledge organization specifically designed for the task of 
constraining linguistic realizations. Given a certain 
application domain, the domain knowledge is mapped 
(classified) into the Upper Model knowledge; in this way, 
for each domain object a proper lexical realization is 
established. As a result, changing the application domain 
requires that only the mapping between the domain and the 
Upper Model knowledge is specified. 
Further developments are connected with the use of 
natural language in a domain which implies an interaction 
with the physical world (as happens in the MAIA system). 
This kind of application will also raise the need to access 
both information gathered from the physical environment 
and dynamically changing knowledge and of a more complex 
pragmatic component, thereby stressing the need for a clear 
architecture. We are also working on the issue of integrating 
such expansions within the approach to multiple underlying 
systems (MUS) as established by \[Bobrow et al., 1990, 
Resnik, 1989\]. In the MUS approach, a user may need to 
combine the capabilities of more than one system (i.e. 
several DBs on various domains, expert systems, 
information retrieval systems, interfaces to simulation 
packages, etc.) in order to perform a general task. For 
dealing with MUS, not only our semantic modules must be 
able to represent various levels of meaning of a sentence, 
they must also be capable, in a transparent manner, of 
organizing the different applications at their disposal and 
choosing which combination of them to use. 

References 

Allgayer, J., Jansen-Winkeln, R., 
Reddig, C., Reithinger, N. "Bidirectional Use of Knowledge 
in the Multi-Modal NL Access System XTRA". In 
Proceedings of IJCAI-89, Detroit, Michigan, 1989. 

Allgayer, J. "SB-ONE + - dealing with sets 
efficiently". In Proceedings of ECAI-90, Stockholm, 
Sweden, 1990. 

Ayuso, D., Donlon, G., MacLaughlin, D., 
Ramshow, L., Resnik, P., Shaked, V., Weischedel, R. "A 
Guide to IRUS-II Application Development". Report No. 
7144, BBN System and Technologies Corporation, 1989. 

Barwise, J., Cooper, R. 
"Generalized quantifiers and natural language". Linguistics 
and Philosophy, 4, 198l. 

Bateman, John A., Kasper, Robert T., 
Moore, Johanna D., Whitney, Richard A. "A General 
Organization of Knowledge for Natural Language 
Processing: tile PENMAN Upper Model". 1SI Technical 
Report, USC/Information Sciences Institute, 1990. 

Bobrow, R., Resnik, P., Weischedel, R. 
"Multiple Underlying Systems: Translating User Requests 
into Programs to Produce Answers". In Proceedings of ACL-90, 1990. 

Boisen, S., Chow, Y., Ingria, R., Roukos, 
S., Scha, R., Stallard, D, Vilain, M. "Integration of Speech 
and Natural Language Final Report". Report No. 6991, BBN 
System and Technologies Corporation, 1989. 

Cooper, R. Quantification and Syntactic 
Theory. Reidel, Dordrecth, 1983. 

Forbes, G. "Indexicals". In: Gabbay & Guenthner 
(eds.), Handbook of Philosophical Logic IV. Reidel, 
Dordrecth 1989. 

Franconi, E. "The YAK (Yet Another Krapfen) 
Manual". IRST Manual, Trento, Italy, 1990. Also as 
'Progetto Finalizzato CNR - Sistemi Informatici e Calcolo 
Parallelo'. 

Hindley, J., Seldin J. Introduction 
to Combinators and &-Calculus. Cambridge University 
Press, 1986. 

Hobbs, J., Shieber, S. "An 
Algorithm for Generating Quantifier Scopings". 
Computational Linguistics, 13, January 1987. 

Keenan, E., Stavi, J. "A semantic 
characterization of natural language determiners". 
Linguistics and Philosophy, 9, 1986. 

Lavelli, A., Magnini, B. "Lexical 
Discrimination within a Multilevel Semantics Approach". 
Proceedings of ALIA-91, Palemlo, Italy, 1991. 

Mac Gregor, R.M., Bates, R. "The 
LOOM Knowledge Representation Language". Techical 
Report ISI/RS-87-188, USC/In formation Science Institute, 
1987. 

Resnik, P. "Access to Multiple Underlying 
Systems in Janus". BBN Report No. 7142, 1989. 

Samek-Lodovici V., 
Strapparava C. "Identifying Noun Phrase References: The 
Topic Module of the AlFresco System". Proceedings of 
ECAI-90, Stockholm, 1990. 

Scha, R. "Logical Foundations for Question 
Answering". Philips Research Laboratories, M.S. 12.331, 
Eindhoven, The Netherlands, 1983. 

Scha, R., Stallard, D. "Multilevel 
Plural and Distributivity". In Proceedings of ACL-88, 1988. 

Stallard, D. "'Answering Questions Posed in 
Intensional Logic. A multilevel semantics approach". BBN 
Report No. 6522, June 1987. 

Stock, O. "Parsing with Flexibility, Dynamic 
Strategies, and Idioms in Mind". Computational 
Linguistics, 15(1): 1-18, 1989. 

Stock, O. "Natural L~mguage and Exploration of 
an Information Space: The ALFresco Interactive System". In 
Proceedings of IJCAI-91, Sydney, Australia, 1991. 

Strapparava, C. "From Scopings to 
Interpretation: The Semantic Interpretation within the 
ALFresco System". In Proceedings of ALIA-91, Palermo, 
Italy, 1991. 

Weischedel, R. M. "A Hybrid Approach to 
Representation in the Janus Natural Language Processor". In 
Proceedings of ACL-89, Vancouver, British Columbia, 
1989. 

Westerstahl, D. "Quantifier in Formal and 
Natural Language". Report No. CSLI-86-55, June 1986. 
