Representation and Interpretation of Determiners in Natural Language

REPRESENTATION AND INTERPRETATION OF DETERMINERS IN NATURAL LANGUAGE

Barbara Di Eugenio' and Leonardo Lesmo

Dipartimento di Informatica - Universita' di Torino C. so Svizzera 185 - 10149 Torino - Italy 39-11-7556771761444

by first order predicate calculus:

"The sentence Floyd broke a glass translates into

3 x (glass (x) " break (Floyd, x)) or-1 in the notation of resmcted quantification ( j x : glass) (break (Floyd, x)).

In either case, the syntactic constituent a glass does not correspond to any semantic constituent; rather, its interpretation forms several discontinuous parts of the logical expression." [Jackendoff 84, p.52].

A principle that should guide the search for a perspicuous semantic theory is the Grammatical Constraint: "The Grammatical Constraint says that one should prefer a semantic theory that explains otherwise arbi­ trary generalizations about the syntax and the lexicon" [Jackendoff 84, p.51]

The first aim of the representation presented in this paper is to respect the Grammatical Constraint, in order to mcrease further the naturalness of the interpretation process. In fact, it will be shown that the analysis can be carried on in a deterministic way. The features of the representation will be described by discussing its suitability to handling a difficult problem in NL analysis, i.e. the interpretation of determiners. The analysis of determiners is strictly con­ cerned with the relationships between sense and reference, intension and extension, concept and class. It is not possible to analyze here in detail the distinction among the different terms *, but we will try to make clear which kinds of linguistic phenomena have to be modelled.

The purpose of a natural language utterance is to make some assertion or to ask for information about some­ thing. This something must be specified by the speaker in some way. To do that in an understandable way, the speaker must make some assumptions about the knowledge available to the hearer. In order to be as informative as possible, and if s/he assumes that the involved entity is known to the hearer, the speaker must try to enable him/her to identify that entity; otherwise, slhe can leave it (more or less) unspecified (see [Appelt 85]). The same identification issues arise if the predicat10n does not concern an individual entity, but applies to each entity of a given set. A very different case IS the one where the predication does not refer to the elements of the set, but to the set as a collective entity: here, the fundamental J'rocess known as "inheritance" does not appl~, so we nee some way to keep apart the two kinds of "set' statements.

Still another different case is the one where there is no specific entity (or set) involved in the predication, but the assertion is "genenc", in the sense that it applies to all possi­ ble entities satisfying the description. In principle, nothing more than context-dependence seems implied by this distinc­ tion; for instance, Exl: "The girls are very mce" expresses a fact about a well-defined set, both in case it is understood

· The most thorough discussion is, to our knowledge, [Lyons 77, chapt.7]. See also [Woods 75] (in particular, the fourth section) and [Castelfranchi , Stock 86].

ABSTRACT.

Following the principles of locality and compositionality during semantic interpretation, we propose a semantic representation formalism which mana~es to deal with refer­ ence problems, i.e. with the interpretation of NPs (in particu­ lar, of those beginning with an article): the addressed prob­ lems are generic I specific I class readings, and collective I distributive interpretations.

The formalism follows the semantic net approach, and uses different representation plans (semantic, content and refer­ ence) and particular structures, called ambiguity spaces, that hide the ambiguities from the other parts of the sentence and remain neutral with respect to the various interpretations until certain disambiguation clues are found.

Besides presenting the formalism, the paper discusses which clues may be used to disambiguate among the various interpretations and shows the way this representation is used to update the hearer's knowledge base *.

I INTRODUCTION

The interpretation of natural language is a process aimed at converting a linear sequence of word forms (a "text") into a semantic representation that can be worked on in order to perform inferences, produce answers, and, more generally, to modify the knowledge base of the hearer **. Many knowledge sources cooperate in a more or less syn­ chronized way to perform the task: they include lexicon, syntax, knowledge about the way utterances are connected to each other, and knowledge on how the structured representa­ tion of a sentence can be converted into the corresponding semantic representation ***. The present paper is concerned with the last of them, and in particular With what is usually called "semantic interpretation process".

One of the most common assumptions about the way this process is carried on is the principle of compositionality, stating that the meaning of a given constituent can be ·obtained by combining in some way the meaning of its com­ ponents, where by "constituent" it IS usually intended a suit­ ably defined portion of the syntactic representation. This principle establishes a clear and effective relationship between the syntactic structures and their semantic counter­ parts (of course the overall context has to be taken into account to solve, for instance, anaphoric references), but it still allows a semantic theory to model the interpretation pro­ cess in an apparently unnatural way. An example is provided as a reference to a subset (contextually determinable) of girls, and in case it is intended to state a generic property of the (context-independent) set of all girls. On the other hand, specific and generic expressions are worth being represented in different ways for two reasons. The first is mainfy seman­ tic: in definite references the set is fixed (i.e. it does not change over time), whilst in generic expressions the property applies to any member of the set, even in case some current members were not in the set when the proposition was issued. The second reason is linguistic: genenc statements can also be expressed in the singular (intuitively, without reference to a set) as in Ex2: "A bear hibernates in winter" or Ex3: "The dog is a friendly animal"*. This is one of the main sources of problems in the design of an effective interpreter of NL expressions.

" Current address: Dipartimento di Automatica e Informatica - Politecnico di Torino - Corso Duca degli Abruzzi 24- 10129 Torino- ITALY.

· This research has been supported by a grant from Ministero della Pubblica ls­ truzione - MPI 40%; it has been carried on while the first author was granted a fellowship from MesarTeam S.P.A. - Torino.

""" We will use the tenns "speaker" and "hearer'' to refer to the producer and the re­ ceiver of the message, without any commitment about the fact that the message is written or spoken.

++· We assume that such a structured representation exists. 1be debate on this as­ sumption is exemplified by [Schank, Birnbaum 84] [Man::us 84] [Tanenhaus et al. 85].

A. Inte~retation problems for determiners

s seen above, the same arttcle may indicate different types of referents. For instance, in case of a definite article, there are three alternatives: a single indivi­ dual (or a specific group of individuals), as in Ex4: "The bear is eating meat"; the generic element of a set, as in ExS: "The bear eats meat" or Ex6: "The bears eat meat" (this sentence is likeky to sound unnnatural to English speakers, because they would say "Bears eat meat"; on the other hand, in Italian we seldom have bare plurals); the class corres.Ponding to a particular concept, as in Ex7: "The bear is dymg out" [Di Eugenio et al. 86]. The choice of the correct mterpretation is not easy and, although some clues may be provided by syntactic, semantic and/or pragmatic knowledge, this choice ts sometimes impossible without tak­ ing into account the discourse history.

Another point from which ambiguity arises is the possibility of having distributive readings, like in Ex8: "The soldiers who hit Hannibal have been praised", or collec­ tive ones, like in Ex9: ''The soldiers who defeated Hanni­ bal have been praised"**. In both cases, the initial representation we give to the NP must be neutral, in order to be specified, if this is possible, during the further processing of the sentence.

Notice that the neutrality required of the initial semantic representation as re~ards the two problems of gen­ eric vs. specific or collective vs. distributive reading is different: m the former case, the ambiguity concerns the interpretation of the NP itself, i.e. what kind of reference is made; in the latter one, it concerns its links with the other parts of the sentence. In fact, a plural NP defines a set through its members, and the set and its individual members must be both present in the representation ***, as either of them may be referred again in the sentence: disambiguating a reading as distributive or collective, then, does not mean to furtherly specify some elements of the representation, but to individuate which element - set or individual member - must be linked to the other parts of the sentence.

Some of the representation formalisms that have been proposed in the literature seem to have a strong expressive power, because they manage to represent the different readings we talked about above (at least some of them), but they lack to explain, or do not stress enough, the way in which this representation is built.

· This is probably one of the reasons that led many researchers to postulate the ex­ istence of "prototypes". Their use and their import have been widely discussed in the litera­ ture (see, for instance, [Brachman 85)); although prototypes are not part of the representa· tion that we propose herein, this representation is amenable to the same discussion. Howev­ er, this is not a paper on knowledge representation, at least in its usual sense, but on seman­ tic interpretation; .this simply means that we will not consider problems about inheritance. cancellation, classifiers and so on, but only the internal structure of the molecular entities which the representation is composed of.

·· Actually, as Webber points out [Webber 83], a plural NP may have another read­ ing, the conjunctive one: "Three boys bought five roses" may mean "Three boys bought five roses (each) ..... Three boys bought five roses (together)", or .. The total of rose-buying boys is three and the total number of roses is five , without any commitment to the relation between each boy and each rose". We will not address this problem, since it seems to be more related to the treatment of quantifiers than to the one of determiners.

··· A similar point is made by Hobbs [Hobbs 83], who states that .. the logical form for a plural NP will include reference to a set and its typical element".

For example, Webber's log!~al formalism, which resorts to restricted quantification l Webber 83], is able to represent definite and indefinite NPs, collective and distribu­ tive readings, generic sets; on the other hand, her aim is to derive entity descriptions that allow her to solve subsequent anaphoras, so that she generates those descriptions from sen­ tences that have already been translated into a logical for­ malism, but the way she represents the NPs appeanng in the sentence while the sentence itself is processed is not clear.

Hobbs's approach [Hobbs 83] is based on a fiat logi­ cal formalism where the predicate SET and the so-called 'typical element' of a set nicely account for distributive and collective readings. The drawback of this representation is that it does not seem to allow the most natural treatment of generic assertions: in fact, the deduction that the individual about which the predication is made is the typical element of a ~e~ is left to the di~course .interpretation process; in our opm10n, ~s the genenc readmg of an NP is sometimes clearly pomted out by clues present in the sentence itself, it is not the case to systematically demand this task to the discourse interpretation process; on the contrary, there must exist a distinct way to represent generic sentences.

. . An?!her flaw of the approaches previously adopted is their mabihty to model the NPs of type 'class'; e.g. in Hobbs' representation they could be solved by using an entity for which the predicate SET holds, but this would fail to dtstinguish between set and class. Notice that assertions on classes are not to be confused with collective readings of sentences, although they are very similar: first of all, 'class' readings are generic assertions, collective ones not neces­ s~lY,; moreover, colle~tive r~adin~s usually express a parti­ :Ipat!on (not necessanly acttve) m some event, something that IS true of every member of the set, while no particination of the single individual to an event * is asserted in class assertions.

A last case, which deserves particular attention, since it is substantially different from all others, occurs in the so­ called Opaque Contexts: no individual is involved, but nei­ ther inheritance applies, nor the class is referred. It is exemplified by sentences as ExlO: "John wants to marry a Norwegian", Exll: "Mary is looking for a knife". In these cases (though these examples are ambiguous) there is a reading where the meaning of tbe indefinite noun phrase can be roughly characterized as "some entity which has the characteristic properties denoted by the term used". Opaque contexts are very strictly correlated with the concept of "intension". What is peculiar is that the 3 quantifier that is usually associated with indefinite singular NPs (denotin& the "existence" in the universe of discourse of an entity satisfy­ ing the predicates that apply to the quantified variable) is loosin~ now its power. In fact, in Ex12: "Every boy needs a dog no existing dog is involved, neither a single one that eyery boy needs, nor a specific <?ne associatecf with each different boy (perhaps via a smtable Skolem function). Instead, what is said IS that for each boy there should be at least one entity to which the "dog" intension applies and which is needed by that boy. Note that it is by no means said that any dog can satisfy the needs of a particular boy, so that the inheritance mechanism cannot apply to dogs, thus inhibiting an interpretation as "a property of each dog is that it is . needed by tli<? boys". It is clear that opaque contexts require a very special treatment.

· The word 'event' here does not mean single occurrence of a particular fact, but it is used in the broader sense of situation, state or action, possibly generic, expressed by a verb phrase.

II TRI-CONCEPTS AND AMBIGUITY SPACES

The solution we propose for the problems outlined in the introduction involves new representation tools that we called Tri-Concepts (TCs) and Ambiguity Spaces (ASs). We introduce them by means of an example.

Fi~. l shows the final representation relative to Ex13: "L orso che vive nella gabbia sta mangiando" (The bear which lives in the cage is eating) *. The representation involves three different plans:

• the semantic ' plan (SP), which contains the semantic knowledge of the system (i.e. the - more or less - static knowledge about the world);
• the content plan (CP), on which the representation of the content conveyed by the sentence is built;
• the reference plan (RP), on which the possible referents of the.NPs are built.

The reference plan, whose existence is limited to the single sentence processing, stores the initially ambiguous represen­ tations of the NPs and not the entities previously mentioned in the discourse; the RP allows to take apart the problems of the referential interpretation of an NP and the semantic role that the object that the NP denotes plays in the sentence. In fact, what seems important is to maintain distinct those features of the representation that concern the semantic roles that constituents play with respect to each other and those regarding the ambiguities mentioned above. We solved this problem by means of the reference and content plans and by means of what we call 'ambiguity sllaces' **: they are struc­ tures that embody the ambiguities mternally, and, in a cer­ tain sense, hide them from the other components of the representation. If the ambiguity can be solved at the syntac­ tic and semantic level, the contents of these spaces are modified to convey this information without affecting the other parts of the sentence; otherwise, the still ambiguous (or partially ambiguous) spaces will be passed to the processes that access the pragmatic knowledge and/or the discourse context.

There are two types of Ambiguity Spaces: the refer­ ence ambiguity spaces (RASs), regarding the genenc I specific I class readings, and the distributivity iunbiguity spaces (DASs), that concern the collective I distributive readings.

Before talking about the ASs, we must explain the KB representation that corresponds to a noun: It is not merely a node, but a ternary structure, called a Tri Concept, that includes three atomic nodes: the concept, the extension and the class.

The concept node will be used in the usual way: it will take part in taxonomies and will accumulate assertions, but only those inheritable by every individual corresponding to that concept.

The c1ass node will accumulate the generic assertions that are not inheritable by individuals (like those contained in Ex7).

The extensional node will not accumulate any assertion, but it will make it possible to have the specific readings of plural NPs linked to the correct set referent and to express membership to a set.

A. Reference ambWui~ sl?aces

When an N hed ts met, a new node on the content plan (x1 in fi~;~ . l) and a RAS on the reference plan are created; the imtial content of the RAS depends on the type of the NP determiner: while the definite determiner may be involved in a generic, specific or class reading (see Ex4, Ex5, Ex6, and Ex7, in section I.A), the indefinite determiner is not used in class assertions: you cannot say "A bear is dying out" or "A bear is studied by naturalists" ***.

· From now on, the examples will be in Italian, because our PIDO system works on the Italian language.

·· Both the three plans and the ambiguity spaces are under some respect similar to spaces in partitioned networks [Hendrix 79].

··· Perhaps the sentences above may be uttered if the meaning of 'a bear' is 'a [type of] bear', but we do not consider this case.

Therefore, the RAS initially associated to a definite NP will contain a complete copy of the TC, the one associated to an indefinite one only the concept and the extensional nodes; besides, the arc connecting the node on the content plan and the RAS will be labelled by SAME for definite NPs, by I­ OF (INSTANCE-OF) for indefinite ones: in the former case the NP refers to a precise individual (or set), concept or class, while in the latter nothing can be said on the indivi­ dual denoted by the indefinite NP besides its belonging to a certain set or its being a random exemplar of the concept (in case of the generic reading of the indefinite article). fn fig. 2 we can see the situation occurring after an NP head has been met (the other parts of the representation are omitted).

This initial representation will be augmented during the processin~ of the sentence if the NP has specifications, e.g. a restrictive relative clause (see Ex13), an adjective (the white bear), a relation with other objects (George's bear): these specifications are not directly linked to the content node that refers to that entity (x1 in fig. 1 ), but to another node (~ in fig. 1) by means of a DEF-AS arc. The meaning of this connection is that the subset, subconcept or subclass that the NP refers to is defined as the portion of the content net indexed by the DEF-AS arc itself. The introduction of a pair of nodes (x1 and x2) is justified by the need of keeping apart the entity that participates in the main predication and tile way it is defined; this is impossible if x1 and x2 coincide.

B. Distributivia ambifiuity s~aces

Io a pural de mteP a second AS, the DAS, is associated on the content plan, as we can see in fig. 2: it allows to represent both the set that a plural NP defines and the members of this set without any commitment to the role that each of them plays in the sentence. Another DAS will be created on the content plan to express the specifications of the NP, in an analo~;~ous way to tnat described above for the specifications of a smgular NP.

In fig.3 the final representation of complex NPs is shown: the RASs are internally transformed according to the particular reading. If the NP has no specifications, it will simply consist of the node, or DAS, connected to the RAS by the arc SAME or I-OF; the RAS will be linked to the TC on the semantic plan by an arc labelled by EQUAL, because no subset, subclass or subconcept has been defined. Note that there is no incompatibility between the representation of an entity on the content plan bY. means of a DAS, i.e. as a pair "set-member" and the posstble class reading that to that r.air can. be given; in fact, we may have sentences like Ex14: 'Gii orsi che vivono in Alaska si stanno estinguendo" (The bears which live in Alaska are dying out), wbere we have a set of bears, which in turn corresl'<>nds to a particular class of bears, upon which a predication ts made.

As far as intensional context are concerned, they are simply represented by marking the RAS with "intensional" and by connecting the CP node with the concept node in the RAS.

III DISAMBIGUATION RULES

Although the task of disambiguating among the different interpretations of an NP is not easy and cannot be fully dealt with without taking into account world knowledge and the discourse history, some useful syntactic and semantic criteria do exist, and they have been easily embodied in the framework constituted by the FIDO system (Flexible Inter­ face for Database O~rations) (see LLesmo Torasso 83], [Lesmo Torasso 85], LDi Eugenio et al. 87]). The features that we have taken into account follow ""·

A. The verb

One useful feature of the verb is its tense: verbal forms different from habitual past and habitual present (and future which needs distinct considerations) expres~ a spe~ific readin'g if the subject is indefinite, while for definite subJects the same is true provided that the verb cannot be referred. to a whole class (like ','estin.guersi", i.e. "~o die <?Ut"). Notice that Italian and English differ 1 because m .English there. are two different forms expressmg prog~ss1ve and habitual present, while in Italian .we hav~ a um_que present form (a progressive form does ex1st, but 1ts use IS not compul~ory as 1t is in English); on the other hand,. we ha_ve two d1ffer~nt simple past forms to express a durat1v,e action, or ~ specific event (iilthough in English the form J.!Sed to ex1s.ts) .. By resorting to the ~erb tense and . some~1mes by takmg 1~to account its meamng, we can d1samb1gu~~e sent~nce~ like ExlS "II leone sta mangiando le mucche .<The lion 1s ell;t­ ing the cows) where we have the progressive form als? m Italian Exl6· ',.I Ieoni uccisero le mucche" (fhe lions k1lled the co~s) E~l7: "II leone si sta estinguendo" (The lion is dying out). A sentence that remains ambiguous is Exl8: "II leone mangia le mucche", that, in Italian, may mean both "The lion is eating the cows" and "The lion eats the cow.s". The future tense gives some problems, as sentences like Exl9: "II leone mangera' le mucche" (The li_on will eat the cows) can be disambiguated only resortmg to world knowledge and the discourse context.

· We will only discuss the problem of generic I specific /class readings, as collcc· tive I distributive interpretations seem solvable by resorting to the semantics of the verb.

B. NP s~ifications

Iefeatures that we have taken into account are a) possessive adjectives and b) restrictive rela~ive cl~us~s.

1. It should not sound strange that possessive. adJectives ml!y help to disambiguate an NP with a determiner, because m Italian possessive adjectives are very <?ften P.recc:ded by a determiner. In most cases, a possessive adJeCtiv~, espe­ cially in the first or second Qerson, makes a ger:enc read­ ing impossible. In Ex20: "II tuo cane mang1a sempre carne" (literally, "The your dog always eats meat"), the specific reading of the subject is f<:>rced by t~e.p~sence of tlie possessive adjective; 10 fact, 1f we orrut It, m Ex21: "II cane mangia sempre carne" (The dog always eats meat) the generic reading is preferred, if the sentence is taken in isolation.
2. Restrictive relative clauses point to a specific reading of the NP they refer to if they contain a reference to a specific individual; this holds both if the relative pronoun is not the subject of the clause, as in Ex22: "The castle that I visited yesterday ·.. ", and if it is the subject and one of the other complements refers to a single individual, like in Ex23: "The dog which bit John is ferocious".

C Mutual influence of NP inte~retations

I here seems to be a certam pre erence for a global generic or s~cific interpretation versus one where some NPs are specific and others .generic. ~n particu.lar, the NP. head ca?­ not be given one Interpretation (specific I genenc) and 1ts specifications the other (generic) specific); the disambigua­ tiOn can go in either direction, i.e. we may disambiguate an NP specification on the basis of the NP head, or the NP head on the basis of one of its specifications. For example, Ex24: "La donna con Ia gonna e' molto ele~ante" (literally, "The woman with the skirt is very elegant' ) may be generic or specific, while Ex25: "La donna con Ia tua gonna e' molto elegante" (literally, "The woman with the your skirt is very elegant") can only be specific. The same considerations hold for different complements of the same verb, although ~e disambiguation of one complement by means of another IS not certain, because counterexamples do exist, like Ex26: "II cane di Giorgio man~;ia le aringhe" (literally, "The dog of Giorgio eats the hemngs"), where, although "the dog of Giorgio" refers to a specific entity, the sentence does not refer to a specific event, but to a habit of that dog.

These disambiguation criteria have been inserted in the FIDO system, a system which accepts natural language sentences and addresses a relational DataBase. The disamoi­ guation criteria have been implemented as condition-action rules; their activation is subordinate to some predefined event, like the analysis of a content word (implemented as a node filling process) or the completi<?n of a constituent (implemented as a node closure .operation). The rule~ are subdivided in packets, each of wh1ch cc;mcerns the defimte. or indefinite determiner and a particular SituatiOn of the parsmg process. The condition parts of the rules. test the already built tree; they examine the NP head, venfy whether there are other disambiguated determiners, check the verbal form etc. The action part augments the internal structure of the RASL · h As an example, .we report a 11;1le relative to t e disambiguation of a defimte NP as a specific reference:

CONDITION:

(AND

ACTION: (CREATE INDIV current)~A.S new_il:~d1v) (LINK current RAS new md1v extensiOn m of) (DIRECT current_ RAS new _indiv content_liiik)

. ) (OR ((CURRENT TENSE present_progress!Ve (CURRENT' TENSE simple _past)) (NOT (CLASS_ VERB current_verb))) .

The condition part verifies that . the cu~nt verb does not require a class subject ~nd that ~ts te~se 1~ present progres­ sive or simple past (r~mm~ that m Italian sim_£le past ~as n,o duration meamng); m th1s case, the specific rea~mg IS accepted and the actions modify th~ Cl.men.t_RAS, 1.e. ~he one connected to the NP under exam10at1on, 10 the follow10g

way: . · d' CREATE INDN creates a new individual node - new _m zv - in the current RAS; .

UNK links new-indiv to the extensional node - extenszon - of the TC contained in the current_ RAS by means of an arc labelled m _of; . .

DIRECT directs the arc content lmk commg. from the con­ tent plan, whose head pointed to the ~AS, onto new _indiv. What is obtained is the structure shown m fig.3 _al.

A last word on the certainty of these rules: while the one above is certain, other rules only give a preferential reading to a given NP (one of these. rulc;:s concerns the mutual influence of two complements, hke m Ex26 above). In case a rule is not certain, its action part is not executed, but it is delayed until other rules are activated that ~ay confirm that interpretation. At the moment, we are study10g a weighting mechanism to balance the interaction of the rules. If contradictions are detected, or if insufficient evi­ dence in favour of an interpretation has been collected, RASs remain ambiguous.

IV HOW THE REPRESENTATION IS USED

The formalism previously described a~ms at representing the information content of a sentence mdepen­ dently of the context where it has been uttered. Of course, the way it is built is just the first phase of the interpretation process; although this paper is mainly concerned with it 1 it seems worth while giving an idea about how the resulting representation is used afterwards.

The described proposal tries to exploit the advantages of a description of the;: interpretation . process. called Transient-Process Account 10 [Woods 75] Without be10g sub­ ject to the same criticism. In other words, the RP and the CP are transient, in the sense that they disappear at the very end, but they are an unambiguous representation of the sentence *. When this representation is made available, the hearer must access his/her contextual knowledge in order to find possible referents or to build structures to associate with newly introduced entities. But now all the information needed for guiding the search has been made explicit. This does not mean that what remains to do is an easy task (cfr. examples like Ex27: "Harry saw an hungry lion; the poor beast ... "), but the network representation makes it rather straightforward to verify that "lions" form a subset of "beasts" (this knowledge should be available on the SP) so that the rules of correct reference can be defined on well­ known subsumption arcs *.

· The discussion in [Woods 75] is primarily addressed to relative clauses, but the distinction we made between "SAME" and "DEF-AS" links allows to keep apart the re­ ferred entity from its description, whatever the fonn of the latter is.

We can now introduce a preliminary version of the rules that govern the execution of the last phase, i.e. the mixing of previous (contextual) and new (RP-CP) informa­ tion.

The updating of the knowledge base of the hearer can be done in two ways, depending on the contents of the Reference Plan (RP) and of the Content Plan (CP).

The contents of the RP determine the mode of this ufdating:

a if there is one or more I (individual) or S (set) nodes on the RP, then the corresponding nodes have to be looked for by the hearer in his knowledge base (to be more precise, in that part of his KB that regards the current situation of the world, analogous to the A-Box in a KRYPTON-like framework [Brachman et al. 85]). The node to be found must satisfy the given description both in terms of the extension of the RP node, as given by the outgoing M-OF or SS arc (i.e., it must be of the right 'type'), and in terms of the possible DEF-AS arcs exiting from the RP node. Moreover, both in case it is an individual and a set, it must be unique, i.e. there must not be two different nodes satisfying the descrip­ tion **. When the relevant node has been found, the assertion about it expressed by the utterance ("ASSERT' entry point) has to be added to the knowledge base.

In case no· node satisfying these requirements is found, two possibilities arise: if the nodes I and/or S are defined by means of a DEF-AS arc, new individuals or new sets are to be added to the KB; otherwise, i.e. if the reference is a bare NP without any specification ("the bear", "the woman"), as we do not consider situa­ tions where the referent is perceptually available, an inconsistency is found: either the speaker has used an inappropriate expression (but this can be disregarded, because we suppose the speaker follows the Conversa­ tional Postulates [Searle 69]), or the system has misin­ terpreted the reference, which is not specific, but either generic or class. At this point, a dtagnostic process should be triggered; our provisional solution is to ask the user.

1. If the reference is definite (arc SAME between CP and RP), but generic or class, the general KB *** will be updated, according to two different modalities: if the NP has no specifications, the assertion will be made on the node (concept or class) in the TC corresponding to the NP head and contained in the KB; otherwise, as there is a DEF-AS arc coming out from the RAS, a new TC will be created, whose concept node will be a sub­ concept of the concept corresponding to the head of the NP (this corresponds to creating a new term in a T­ Box) and the assertion will be stored there ****.
2. In case the reference to a single individual is indefinite (arc I-OF pointing to an EXT node), then an individual node corresponding to the description (in the same

· Of course, if Harry was my dog and it was devoured by the lion, the reference as­ sumed in the text is not the right one. Moreover, it is clear that any mechanism for access­ ing the previous context should pay attention to more general problems of anaphoric refer­ ences, and in particular to their relationship to the organization of discourse - on the prob­ lem of focussing see, for ex., [Sidner 83], [Grosz 81 ].

·· This search process has not yet been implemented. Of course, it is not only a matter of efficiency, but also of taking into account the relevant data. For instance, the uniqueness condition must apply to a suitably specified current context, othetwise "the boy" or "John" could never be used.

··· Its contents are similar to aT-Box's [Brachman et at. 85].

···· Notice that we do not consider really definitional sentences like "The whale is a mammal".

sense explained above) must be created in the knowledge base.

d) In case an indefinite reference gives rise to a generic reading (arc I-OF pointing to a CONCEPT node in a RAS), this generic assertion is stored on the concept node; how­ ever, the problem is still under investigation, because this use of the indefinite article does not imply generic universal­ ity (an assertion of this kind can be considered 'defeasible' [Croft 85]).

e) The final point regards intensional contexts. The problem of their interpretation (i.e. what to do when the RAS is marked as "intensional") is still open. The solution we are currently implementing is to maintain nested spaces which block some kind of inferences. For instance, in the case of Ex12 (Every boy needs a dog), we get something that correspond to a formula as:

where the higher order predicate NEED is opaque in the usual sense. As the paper is mainly concerned with the pre­ vious phase of the interpretation and the import of this solu­ tion is currently under study, we will not go in more details here.

A last word on the situation in which the RASs on the RP are ambiguous: the disambiguation task is dealt with by the processes that access world knowledge (however, they take into account the possible clues in favour of a certain interpretation found during the analysis of the sentence). These processes act as a filter that eliminates some interpre­ tations; e.g. in Ex26: "L'uomo ha sofferto per molti secoli" (literally "The man has been suffering for many cen­ turies") the maximum length of a single person's life forces the generic interpretation.

V CONCLUSIONS

This paper introduces a new formalism for representing the information contents of a NL utterance. The main advantages of this rroposal are:

• the different readings of an NP can be represented in an unambiguous way. That is, distinct representations are obtained for specific, generic, class, collective, distribu­ tive, intensional readings;
• ambiguous expressions result in partially · specified representations; in principle, other KSs could operate on them in order to solve ambiguities;
• the analysis can be carried on in a deterministic way, i.e. without reshaping temporary structures, but only adding new data.

The whole interpretation process is assumed to work in two steps that are currently applied sequentially, but could be interleaved in a subsequent version. The first step, with which this paper is more concerned, regards the construction of a context-independent representation of the sentence. The second one regards the actual update of the hearer's knowledge base on the basis of the representation that has been built. In this second step context-dependencies are taken into account.

We believe that the approach described in the paper is a first step towards a p<?rspicuous treatment of quantifiers in natural language, whtch accounts both for the intrinsic complexity of the matter and for the apparent ease that seems to characterize human processing of q_uantified phrases. The paper does not treat generalized quantifiers but only determiners; the results obtained for them are encourag­ ing and the studies on real quantifiers [Di Eugenio, Lesmo 86J seem to show that the formalism can be extended quite easily to cover them.