next up previous contents
Next: Lexical Semantic Relations Up: Linguistic aspects of lexical Previous: Introduction


Lexical aspect


The notion of aspectuality has traditionally come to cover a number of distinct but closely related phenomena and approaches strictly intertwined with the notion of event and event structure. Consider the following examples:

(1) a. ? John loved Mary for/in three years
b. John ran for/*in three hours
c. John ate an apple *for/in three minutes
d. John reached the top of the mountain *for/in three days
As can be seen, the reported verbs differ as to their capability to combine with for- and in-adverbials. This and many other examples lead scholars to revive and renew a classification of verbal predicates dating as far back as Aristotle. For instance, it is now common to talk about stative predicates (to love, to know, ...), activities (to run, to push, ..), accomplishments (to eat, to drink, to write, ...) and achievements (to find, to note, ...), this way following suggestions elaborated by [Ken63], [Ven67], [Dow79] and many others.

It has been noted that these predicates differ as to their internal constitution and their temporal properties. Consider, for instance, (1a). There is a clear indication of a three-year period in which the predicate (to love) was true. However, it can be noted that much the same obtains for any subperiod of it. That is, for any subperiod t of the three-year period, the predicate to love can still truthfully apply. Similar considerations hold, at least to a certain extent, of activities. Surely, there are subperiods of a running that still qualify as running; that is, they show the subinterval property. On the other hand, predicates such as eating an apple do not display the same behaviour. If it took three minutes to eat an apple, as in (1c), then there is no subpart of this period to which the predicate eat an apple can be applied. In this respect achievements pattern with accomplishments. In the literature, these facts have been described by resorting to the notion of cumulativity and/or divisibility. Statives and activities are divisive whereas accomplishments and achievements are not. Other properties have been discovered, which are similar in nature to the one just discussed, e.g. summativity and so on. 2.2

The explanation of these facts crucially depends on a previous understanding of the ontological and semantic features of the involved entities, that is events. In this respect, two main approaches deserve consideration, the Montagovian one and the Davidsonian one.

According to the Montagovian tradition, events are properties of time, whereas Davidsonians regard them as primitive individuals in the ontology of natural languages. For a Montagovian, actionality properties are crucially explained by resorting to time entities. Thus, verbal predicates are true at time points or intervals, and a stative verb corresponds to a predicate which, whenever true at a time (interval) t, is also true at every subinterval included in t. All the other properties can be similarly reduced to properties of properties of times. For our purposes, it is important to notice that this way of explaining events require the full blown apparatus of higher order logic typical of the Montagovian tradition. This means that the approach might be of little value (at least in the majority of cases) for NLP, because of the well known problems of implementing such logics.

In the Davidsonian tradition, as elaborated by, e.g. [Hig85], [Hig97], [Kri89], [Kam93], not only events are primitive individuals, but event variables are introduced in the logical forms of sentences by verbs, nouns (as, e.g. in eventive nominalization) and so on. This paves the way for a rich, and relatively simple (basically first order) semantics of events, in which eventive variables can be quantified over by such devices as adverb of quantification (often, never, always, ecc.., see [Lew75], [Swa91]), modifiers directly attached to events (e.g., by conjunctive modification, as in [Dav67], [Hig85], [Hig97], and so on). In this respect, and partially independently from theoretical reasons, the Davidsonian approach recommends itself for NLP purposes, exactly for the reasons that made Montagovian approaches indigestible for NLP.

Turning to actionality-like properties in a Davidsonian framework, they can be explained by directly resorting to properties of events, leaving time in the background. One way of pursuing this approach consists in singling out some structural primitive, e.g. part-of, which gives rise to appropriate algebrical strucures (lattices and/or semi-lattices). The properties discussed above can then be seen as properties of predicates defined on these algebraic structures. This approach has been explored and made popular by [Kri89], who follows similar suggestions for the objectual domain advanced by [Lin83] and by previous work by [Bac86].2.1 One main point of Krifka's proposal concerns the existence of formal relations between objectual and eventive domains (both conceived as algebraic structures) which take the form of homeomorphisms. These relationships are taken to be responsible for such contrasts as that in (2):
(2) a. John ate an apple in/*for three minutes
b. John ate apples *in/for three minutes
It has been noticed that the nature and properties of the direct object of such verbs as to eat yield different acceptability results when confronted with in- and it for-adverbials. Verkuyl (1993) observes that the existence of a specific quantity of matter as a referent for the direct object can affect the telic nature of a predicate. According to Krifka, this is correct and amounts to a passage of formal properties, via homomorphism, from the objectual domain to the eventive one. The nominal an apple in (2) identifies a specific quantity of a certain matter (+SQA for Verkuyl, a quantised predicate for Krifka). Via the homeomorphism, this induces a similar specification on the eventive predicate, which becomes quantised. On the other hand, given that the bare plural apples is not quantised, by the same token the corresponding complex predicate eat apples is not quantised as well. Thus, the contrast is explained by stipulating: that in- for-adverbials select for quantised/non-quantised eventive predicate; that there can be a passage of formal properties from arguments to predicates, via homeomorphisms; and that such a property is encoded by thematic-relations.

Besides the merit of this, or other, particular theory, the relevant point is that actionality emerges not as a property of isolated verbs (or verbal stems) but, rather, as a compositional facet of complex structures. For our purposes, the morale is that verbal stems need to bear appropriate specification that can be used by a parser/generator and/or by a semantic processor to compute the correct actional value of a complex phrase, taking into account the properties of designated arguments.

The latter point needs further clarification. In (2) the argument affecting the actional properties of the (complex) predicate was the direct object. In other cases, however, other complements/arguments can play a similar role. This is the case of directional PPs with verbs of movement:
(3) a. John ran for/*in two hours
b. John ran home *for/in two hours
These cases are basically similar to those discussed above, with the difference that the affecting factor is not a specified quantity of matter, as with many accomplishments, but a spatial path (in the sense of Jackendoff) provided with a specific end point. Notice the following example:
(4) John walked around the fountain for/in ten minutes.
Both the in- and the for- adverbials are acceptable. However, the sentence changes its meaning accordingly. If the for-adverbial is chosen, then no completion of the around-the-fountain tour is entailed. That is, John might have been walking around the fountain for a certain period of time, without completing its tour. On the other hand, the in-adverbial forces a completive (telic) reading. Although the explanation of this contrast calls for such notions as telicity and perfectivity which will be discussed in the next section, it is important to notice that the contrast is consistent with the discussion above. The PP around the fountain can specify the end point of a path and it does so only with the in-adverbial and not with the for-adverbial.

Perfectivity and telicity

As can be seen from the discussion above, actionality is the result of the interaction between lexical properties (basically, aktionsarten of the verb) and structural ones (the characteristics of the designated argument/complement). Aspectuality, in the way this term is mostly used in the literature, is a morphosyntactic property, deriving from specifications borne by certain verbal forms, such as the Italian simple past, or from free/bound morphemes that can (more or less) freely combine with the verb, as in Slavonic and many other languages.2.3 The basic distinction is that between perfectivity and imperfectivity. Interpretively, perfective verbal forms refer to complete/finished events, as in:
(4) John ate an apple.
The eating, as reported in (4), is finished, completed. On the other hand, imperfective forms lack this meaning and depict ongoing, unfinished processes:
(5) Mario mangia/mangiava una mela.
Mario eats/ate(IMPF) an apple.
The present and the imperfect tense in Italian are both imperfective. As a consequence, a sentence such as (5) does not say anything about the completion/termination of the event described. Even if the imperfect, is a past tense, the event need not be conceptualised as finished:
(6) Alle tre Mario scriveva la lettera e la sta ancora scrivendo.
At three, M. wrote(IMPF) the letter and he is still writing it.
Aspectuality heavily interacts with actionality. Thus, the aforementioned tests with in- and for-adverbials only make sense with perfective forms:
(7) Mario scriveva una lettera *per/*in tre ore.
M. wrote(IMPF) a letter for/in three hours.
Furthermore, the perfective/imperfective distinction also plays a crucial role for telic readings. Teloses (also called culminations, natural end points, etc..) only arise with perfective aspect and can be revealed by the in- for-adverbials test.

Also, teloses depend on the actional value of the complex predicate. Thus, such predicates as eat an apple yield teloses, if perfective. On the other hand, predicates such as eat apples, run, etc.. remain non-telic even when perfective. In a sense, and for many purposes, the telic-atelic distinction overrides and encompasses many traditional aktionsarten ones, along the following schema:
(8) a. Process & +SQA = telic
b. Process & -SQA = atelic
Here process is to be understood as a term of art encompassing vendlerian activities and accomplishments. Notice, that, according to what was said above, the fact that a predicate is telic, in the sense just defined, does not entail that any actual telos is given, because aspectuality must be considered:
(9) a. telicity & perfectivity $\rightarrow$ +telos
b. telicity & imperfectivity $\rightarrow$ -telos
Atelic predicates, on the other hand, never give raise to teloses:
(10) atelic $\rightarrow$ -telos
Finally, there are predicates which always give produce teloses, that is achievements. In this respect, they can be seen as being both lexically telic and perfective.

To sum up. Actionality results from the combination of the lexical properties of the verb together with those of a designated argument. Aspectuality, on the other hand, is a morphosyntactic properties, depending on the presence of appropriate specifications bore by particular verbal forms or by bound and/or free morphemes attaching to the verb. Finally, telicity lies somehow in between actionality and perfectivity, in many respect depending on both. However, we saw that there is a class of verbs, the so called achievements, which are always telic and always actualise teloses.

Before concluding this sketchy introduction to actionality and aspectuality, one point must be stressed. First of all, actionality seems to be a cross-linguistically rather stable property. Thus corresponding verbs in different languages tend to have the same actional properties and much the same can be said about complex actionality. On the other hand, the perfective/imperfective distinction is highly language-dependent. This holds both with respect to the means available for expressing the distinction itself (complex verbal forms, as in Italian, French, etc., vs. free or bound morphemes attaching to the verb, as in Slavonic, etc.) and with respect to the range of possibilities available. Thus, [Gio97] argue that English does not have the perfective/imperfective distinction. In this language all eventive verbal forms are perfective. Similar conclusions can be drawn for many non-indoeuropean languages (Creole languages, African languages, Chinese, etc.).

Overlapping area and phenomena

In this section we will very briefly consider areas and phenomena which overlap those discussed above. The first point to be addressed is the progressive form (found in English, Italian, and, to a certain extent, in other Romance and Germanic languages). Two different perspectives can be taken. According to the first (see Giorgi and Pianesi, 1997) the progressive should be distinguished from true imperfective forms. The former, in fact, has an intensional meaning that ordinary imperfective forms lack.2.4 The progressive should be regarded as a way to imperfectivise an otherwise perfective form by intensionalising the completeness/finiteness of the event. The distinction is subtle, but very clear empirically:
(11)John is reaching the top of the mountain.
Gianni sta raggiungendo la cima della montagna.
*Gianni raggiunge la cima della montagna.
Both in Italian and in English, an achievement predicate such as reach the top/raggiungere la cima can be progressivised in the present tense. However, in Italian, the imperfective present tense form is not acceptable.

On the other hand, other scholars more or less tacitly assume that there is no real distinction between progressive and imperfective verbal forms, collapsing them under the same heading of imperfectivity.

As said above, the distinction is subtle and can probably be neglected for many purposes. It cannot be ignored with such tasks as text or dialogue generation and machine translation, though, at pain of producing a wrong or ungrammatical form. So care must be taken in the design of the basic aspectual classes for an NLP system. Another important overlapping area is constituted by all such aspectual forms and values as inchoatives, resultatives and so on. It should be noticed that many of the deviant examples discussed above can have one of this readings. Thus (3a) becomes acceptable in following context:

(12) After the car accident, John ran in three days.

(12) has the inchoative meaning that John was able to run again only three days after the accident he suffered. We suggest that these cases be left unaddressed within this document. To be sure, the availability of such readings depends on a number of structural (syntactic) facts that are beyond the scope of our contribution.2.5 As such they should be treated by the syntactic and/or by the semantic analyser. In the case such aspectual values are borne by lexical items (aspectual verbs) or by free/bound morpheme, specifications similar to those we are going to propose for the basic aspectual opposition can be easily devised.

Relation to other Areas of Lexical Semantics

Most obviously, lexical aspect and actionality is connected with Semantic (or Thematic) Roles 2.4. For instance, within the approach advocated by [Kri89], the transfer of formal properties from the objectual domain to the eventive one depends on the nature of thematic relation involved, as analysed according to the suggestion in [Dow89] and [Dow91]. See 2.4 for a survey of the relevant problems.

Encoding in Lexical Databases

Attempts have been made at capturing lexical aspect and aktionsarten within LDBs. An example of a static definition can be found in DIONYSUS [Nir93a] where aspectual specifications are provided which distinguish between phasal properties (e.g., begin), durational properties ($\pm$ prolonged), and so on.

In ACQUILEX an attempt is made at a more dynamic classification of aktionsarten which also tries to address the interactions with thematic roles, [San91]. All this is done within a type feature structure (TFS) formalism. Thus, the traditional vendlerian distinction of verbal aktionsarten is reconstructed in the type hierarchy by means of the two types stative and dynamic. These are then combined with the two krifkean types cumulative and quantized to obtain a quadripartite classification of verbal predicates. Such a classification is then connected with the properties of thematic (proto-)roles to allow for the transfer of properties from the objectual to the eventive domain.

Relevance to LE Applications

Concerning NL analysis in general, lexical aspect and actionality play a role in determining the grammatical and semantical interactions between predicates and other parts of the sentence, both arguments and modifiers. With respect to the former, besides the facts discussed in the previous sections, we want to mention genericity, habituality, event quantification, the mass/count distinction and so on, as cases in which the aspectual and actional properties of the predicate play a relevant role in determining syntactic and semantic facts. Concerning adjuncts, the actional properties of a (simple or complex) predicate determine the kind of temporal modifiers it can combine with (for- vs. in- adverbials), the kind of adjunct clauses (purpose and rationale clauses), etc. For all these reasons, lexical aspect and actionality play an important role in parsing (e.g., to restrict and control attachment ambiguities) and in semantic computations.

For the purposes of NL generation, and independently of the framework chosen to express the meaning of the text to be generated, the aspectual and actional properties of lexical items play a crucial role in lexical choice and in microplanning 4.4, These considerations are even the more important when the system deals with multilingual generation. In such cases, in fact, stipulation of direct correspondences between actional and aspectual configurations and predicates on the basis of a single language inevitably fails to generalise to other languages, so that more principled approaches are necessary.

Concerning classes of applications, it is worth mentioning that because of the kind of (semantic) information provided, lexical aspect is important for all systems which must be able to detect and reason about events. This is true of information and knowledge extraction systems 4.2, dialogue management systems, etc... and in general for all the cases in which it is important to tell whether an action/event, as presented in the text or during the dialogue, is terminated/finished or whether it can continue at the speech time or at any other focalisation time. Furthermore, actionality is crucial to establish the appropriate relationships between the various events presented in the text/dialogue: simultaneity, overlapping, precedence, etc... Within machine translation 4.1, given the various means available to different langauges to express actionality and aspect, a proper mastering of these notions is crucial to improve the quality of the translation, [Dor93], by selecting/producing the appropriate constructions in the target language, etc., irrespective of the chosen approach (interlingua, transfer, mixed techniques based). Finally, we want to emphasise that the strict relationship between the eventual domain and the temporal one makes the aspectual information indispensable for systems performing any form of temporal reasoning: e.g., task scheduling, dialogue management, population and maintainance of temporal databases, etc..

next up previous contents
Next: Lexical Semantic Relations Up: Linguistic aspects of lexical Previous: Introduction
EAGLES Central Secretariat