A Generative Perspective on Verb 
Alternations 
Manfred Stede* 
Technische Universit/it Berlin 
Verb alternations have been researched extensively in linguistics, but they have not yet received 
a systematic treatment in natural language generation systems; consequently, generators cannot 
make informed choices among alternatives. As a step towards overcoming this discrepancy, we 
review some linguistic work on several prominent alternations, revise and extend it, and suggest 
a set of rules that allow the series of alternated forms to be produced from a single base form of the 
verb, the lexical entry. The framework has been implemented in the Moose sentence generator, 
which can thus choose a particular verb alternation in order to accomplish generation goals such 
as placing emphasis on the most important element of the sentence. 
1. Introduction 
In this paper, we approach the problem of verb alternations from the perspective of 
knowledge-based natural language generation (NLG), which aims at producing a set of 
verbalizations from a common underlying representation. This viewpoint places some 
specific requirements on the nature of lexical representations, which will be explained 
in Section 2. Thereafter, we will investigate the problem of systematically generating 
a number of verb alternations--a problem that so far has received little attention in 
the NLG community. 
Why should a language generator have knowledge about producing alternations? 
Because a sophisticated discourse production may call for using a verb in one or an- 
other configuration, depending on the current situation of utterance. Alternations can 
place the emphasis on different elements of the sentence, and distribution of empha- 
sis is influenced by the development of the discourse--and thus related to a whole 
range of other generation decisions. To illustrate, consider the following alternative 
beginnings of a little story: 
(1) (a) Tom was in a hurry, but he had to change the oil before hitting the 
road. He crawled under the car and unscrewed the drain bolt. The 
engine drained in 20 seconds .... 
(b) Time was short, but the oil had to be changed before Tom could hit 
the road. Within 20 seconds, he drained the engine. Then ... 
(c) Tom was in a hurry, but he had to change the oil before hitting the 
road. Quickly, he crawled under the car and unscrewed the drain bolt. 
For 20 long seconds, the oil drained from the engine .... 
* TU Berlin, FB Informafik, Sekr. FR 6--10, Franklinstr. 28/29, 10587 Berlin, Germany 
(~) 1998 Association for Computational Linguistics 
Computational Linguistics Volume 24, Number 3 
(d) Tom was in a hurry, but he had to change the oil before hitting the 
road. Crawling under the car, he drained the old oil from the engine, and 
then... 
Depending on how the story develops, which might hinge on stylistic parameters, a 
different configuration of to drain, with different subjects, objects, and prepositional 
phrases, should be produced. Therefore, if an NLG system is expected to be able to 
cope with such differences, it needs knowledge of what alternations are possible for 
a given verb, and how the different syntactic configurations relate to differences in 
meaning. 
In linguistics, the central goal of research on alternations is to uncover the relation- 
ships between syntax and semantics (linking rules), and to form classifications of verbs 
according to their alternation behavior (Levin 1993). To accomplish these goals, the 
need for fine-grained lexical-semantic representations is pointed out, although there is 
no strong consensus yet on exactly what such representations should look like (see the 
discussion in Levin and Rappaport Hovav \[1995, chapter 1\]). NLG, in any case, needs 
representations to work with; and in order to account for verb alternations, we need 
to devise rather fine-grained ones. In particular, a generator has to relate the (possible) 
changes in meaning to the changes in form, so that--from a given representation--the 
correct set of "alternated verb forms" can be produced. In other words, the generator 
needs to know the conditions under which some input representation licenses the use 
of a specific alternation. 
As a step in this direction, we consider a number of alternations that affect the 
aspectual category (or Aktionsart) of the verb--a group that Levin (1993, 12) chose 
not to focus on. We look at the differences in meaning that coincide with such alter- 
nations, and propose suitable representations for input specifications (the underlying 
ontology), for verb meaning, and for the alternation rules. These rules are able to 
sequentially derive the various alternated forms from a single base form, which is 
stated in the lexical entry. The approach has been implemented in a bilingual genera- 
tion system, which can produce the "alternated paraphrases" in English and German. 
To demonstrate its capabilities, we will show how a salience parameter associated 
with the input can give rise to selecting one or another of the alternatives; specifically, 
the production of the alternative drain sentences (1a-d), from a common underlying 
representation, will be demonstrated. 
The paper is organized as follows: Section 2 discusses the specific requirements of 
lexical information in NLG, focusing on verbs, and suggests a format for dictionary 
entries. Section 3 develops our approach to verb alternations and proposes rules that 
account for several such alternations. Section 4 describes the implementation of the 
alternation framework in the Moose generator, and Section 5 concludes and compares 
our approach to related research. While Sections 2 and 3 are concerned with linguistic 
representations and make only little reference to NLG, Section 4 presupposes some 
knowledge of the generation concepts and systems that Moose is built upon. 
2. Representing Verb Meaning for Generation 
2.1 Lexical Information 
The central characteristic of knowledge-based NLG is the existence of a domain model, 
which anchors the representations serving as input to the generator. The domain model 
provides the basis for drawing inferences on the representations. One class of such 
inferences is the subsumption check, which our approach exploits for the task of 
language generation (see Section 4). The NLG system is thus in charge of mapping a 
402 
Stede Verb Alternations 
network of instantiated domain concepts to linguistic utterances, and the specific role 
of lexemes within this process is to "carry over" the meaning from the underlying 
domain model to utterances in natural language. To accomplish this step, the lexicon of 
the generator requires two basic components: it has to explain what words mean with 
respect to the domain model and it has to explain how words can be combined into 
well-formed sentences. At this point, the study of lexical semantics becomes relevant-- 
it should systematically relate these two tasks. 
In NLG, however, lexical semantics has for a long time been relatively neglected; 
words and concepts were often conveniently put into a simple one-to-one correspon- 
dence. The key to incorporating lexical semantics into NLG is breaking up this tight 
correspondence and arranging for more flexible mappings. As soon as entities in the 
knowledge representation scheme do not correspond to words in a direct manner, the 
relationship between word meaning and entities in the knowledge base (KB) needs 
to be specified in some more elaborate way. Now, lexical semantics has to supply the 
interface between knowledge and words: It has to specify what words can be used to 
express what parts of what KB entities, and, possible under what circumstances. To 
this end, the relevant work in linguistics needs to be identified, extended, and adapted 
for generation purposes. This adaptation is not straightforward, however, because the 
starting position of linguistics is different from that of NLG: For a linguist, the syntax- 
semantics interface is of central concern, whereas in NLG, there is the additional level 
of instantiated knowledge, which needs to be systematically related to linguistic levels. 
Our approach is to employ a level of sentence-semantic representation that mediates 
between the knowledge-level input and syntactic realization. A well-motivated se- 
mantic level allows us to encapsulate all syntactic decisions in a front-end generation 
module; we use the Penman system (Penman Group 1989) for this purpose. For the 
other mapping, from KB to sentence-semantics, we use the lexicon as the primary 
source of information. 
In this approach, lexical entries therefore have two main components: a denota- 
tion that defines the applicability conditions of the lexeme with respect to the domain 
model (i.e., it can be matched against the generator's input), and a partial semantic 
specification (PSemSpec), which specifies the contribution that the lexeme makes to 
sentence meaning. 1 The task of lexicalization in our generator thus consists of first find- 
ing lexemes that can convey parts of the input, and then determining the preferred 
combination of candidate lexemes, yielding a sentence-semantic specification (Sem- 
Spec), which is then processed by the surface generator. The generation architecture 
is presented in Section 4. 
The next two sections explain the denotation and the partial semantic specification 
associated with verb lexicon entries, and thereby also the two levels of representation 
used in the generation system. 
2.2 Event Ontology and Aktionsart 
The development of the domain model and the underlying ontology for our system 
focused on the treatment of events so that they can be appropriately verbalized in 
different languages. The hierarchy of situations, shown in Figure 1, is organized along 
a variant of the ontological categories proposed by Vendler (1967) and developed 
further by Bach (1986), inter alia. We briefly discuss the three types of situation in 
turn. 
1 The lexical entries in our system have several other components, which are listed in Section 2.4. 
403 
Computational Linguistics Volume 24, Number 3 
SITUATION 
STATE ACTIVITY /\ 
PROTRACTED- MOMENT.- 
ACTIVITY ACTIVITY 
Figure 1 
EVENT 
CULMINATION TRANSITION 
PROTRACTED- MOMENT.- 
CULMINATION CULMINATION 
Situation types in the ontology of MOOSE. 
States are seen much in the same way as Bach sees them: Something is attributed to 
an object for some period of time, and the object is not perceived as "doing" anything. 
The bottle is empty is true for the bottle without it doing anything about it. We do not 
make further distinctions among states here. 
Activities were called "processes" by Bach, but we need this term on a different 
level of description (see below). They are quite similar to states, but there is always 
something "going on," as in The water wasflowing toward the sea. We distinguish two sub- 
types here: protracted activities take place over an extended period of time, whereas 
momentaneous activities occur in an instant; a "point adverbial" such as at noon serves 
as a linguistic test. 
Events are occurrences that have a structure to them; in particular, their result, or 
their coming to an end is included in them: to destroy a building, to write a book. As their 
central feature we take them to always involve some change of state: the building loses 
its integrity, the book comes into existence, or gets finished. While Bach (1986) did not 
investigate the internal structure of events, others suggested that this needs to be done 
(e.g., Moens and Steedman 1988; Parsons 1990). Pustejovsky (1991) treated Vendlerian 
accomplishments and achievements as transitions from a state Q(y) to NOT-Q(y), and 
suggested that accomplishments in addition have an intrinsic agent performing an 
activity that brings about the change of state. 
We follow this line, but modify it in some ways. Basically, we see any event as 
involving a change of state; an activity responsible for the change can optionally be 
present. A plain transition is necessarily momentaneous (The room lit up), whereas 
a transition-with-activity inherits its protracted/momentaneous feature from the em- 
bedded activity. We call these tripartite events culminations. 2 They are composed of a 
pre-state (holding before the event commences), a post-state (holding when the event 
is over), and an optional activity that brings the transition about. Generalizing from 
Pustejovsky's proposal, we take state transitions to be more than merely oppositions 
of Q(y) and NOT-Q(y); they can also amount to a gradual change on some scale, or 
involve other values. Also in contrast to Pustejovsky, we do not regard the presence 
of a volitional agent as responsible for any of the category distinctions; rather, the 
agentivity feature cuts across the categories discussed. Other aspects of our ontology 
are designed following proposals by Jackendoff (1990), in particular his analysis of 
movement events. 
2 Moens and Steedman (1988) also use this term, but they restrict it to momentaneous events. 
Unfortunately, the terminology used in the literature for these kinds of categories varies so much that a 
standardization seems out of reach. 
404 
Stede Verb Alternations 
event- 1  fill 
con~f----~ > 'not-full 
-state-1 ~ ' '~ 
~f~ > pa~>b~tination ~ 
fill-state-2 ...... ;~.~ ~ water-1 
\] value\[ > 'full 
Figure 2 
SitSpec representing a fill-event. 
Subsumed by the general ontological system, a domain model is defined that holds 
the concepts relevant for representing situations and that specifies the exact conditions 
for their well-formedness. We use the term SitSpec for a network of instances of 
domain model concepts, which will be the input to our generator. The root node of 
any SitSpec is of the type situation. As an example, the event of a person named 
Jill filling a tank with water is shown in Figure 2 in a graphical KL-ONE notation 
(Brachman and Schmolze 1985), with relation names appearing in boxes. The event 
combines the activity of Jill pouring water into the tank with the fill-state of the tank 
changing to full. A verbalization of this event can emphasize either of these aspects. 
Since we decompose event structure in such a way, it follows that the denotations 
of verbs for verbalizing events need to be fairly complex. The type of event denoted 
relates to the Aktionsart of the verb: the inherent features characterizing (primarily) the 
temporal distribution of the event denoted. 3 A generator needs to know these features 
when verbalizing different kinds of events, so that it can produce (for example) the 
correct temporal modifier to express the duration of either an activity or a culmination. 
The variety of phenomena in Aktionsart are far from clear-cut, and there is no generally 
accepted and well-defined set of features. In the following, we use the terms given by 
Bussmann (1983) and discuss only those Aktionsart features that are directly relevant 
for us because they relate types of situations to denotations of verbs. Thus, within 
the context of our system, we define Aktionsart features in terms of patterns of verb 
denotations. Table 1 lists the correspondences. 
Simple cases are stative verbs like to own or to know. Durative verbs characterize 
continuous occurrences that do not have internal structure, like to sleep, to sit. In the class 
of nondurative verbs we find the semelfactive ones, which denote a single occurrence, 
thus in our system a momentaneous activity, as, for example, to knock. Interestingly, 
an iterative reading can be enforced on a semelfactive verb by a durative adverbial: 
She poked me for an hour. Transformative verbs involve a change of some state, without 
a clearly recognizable event that would be responsible for it: The room lit up. The 
denotation of such verbs thus involves a pre-state and a post-state. In our ontology, 
these are transitions. Resultative verbs, on the other hand, characterize situations in 
3 This is often treated on a par with aspect, but we prefer to make a terminological distinction between 
the grammaticalized categories such as progressive versus nonprogressive in English (aspeCt),. and the 
static verb-inherent features. 
405 
Computational Linguistics Volume 24, Number 3 
Table 1 
Correspondences between Aktionsart and 
denotations. 
Aktionsart Denotation pattern 
stative (state X) 
durative (protracted-activity X) 
semelfactive (moment aneous-act ±vit y X) 
transformative (event (PRE-STATE X) 
(POST-STATE not-X) ) 
resultative (event (ACTIVITY X) 
(POST-STATE Y) 
causative (activity (CAUSER X)) 
which something is going on and then comes to an end, thereby resulting in some 
new state (culminations in our ontology). Their denotation includes an activity and 
a post-state. In the literature, such verbs are often also called inchoative. 4 The final 
verb-inherent feature we use is the well-known causative, which reflects the presence 
of a causer in the denotation (as in Figure 2). 
Verb alternations, as we will discuss them shortly, can involve a shift in Aktionsart 
and thus a systematic change of the denotation. But first we have to introduce the 
second major component of verb semantics--sentence meaning. 
2.3 Sentence Meaning 
A SitSpec representing a possibly complex event structure can be verbalized by a va- 
riety of sentences, which can differ in terms of their argument structure, aspectual 
composition, etc. From the viewpoint of NLG, we wish to select the most appropriate 
sentence on the grounds of target parameters, such as the salience assignment men- 
tioned in the beginning of the paper. In order to produce a sentence that accomplishes 
semantic goals of this kind, it is impractical to map the very abstract SitSpec directly 
to a syntactic structure. Instead, we use a sentence-semantic level of description that 
allows us, on the one hand, to control those generation decisions that affect the mean- 
ing of the sentence, and, on the other hand, to encapsulate the syntactic realization 
decisions in the front-end generation grammar. 
2.3.1 Halliday's Ideational Structure. To describe sentence meaning, we use the "ide- 
ational structure" introduced by Halliday (1985). It resembles other approaches based 
on semantic case roles, but an important feature of Halliday's work is his thorough 
classification of process types and of the semantic relationships holding between the 
verb and the other elements in a clause. 5 This extensive analysis renders the approach 
particularly useful for sentence generation. Halliday's process classification has been 
further developed for NLG purposes by C. Matthiessen, J. Bateman and others (see, for 
instance, Matthiessen and Bateman \[1991\]). The resulting "upper model" (UM) is part 
of the Penman generator and used in our system as well. The UM is a taxonomy of 
4 The term inchoative is used to cover a rather broad range of phenomena, including the beginning of an 
event (e.g., to inflame) or its coming to an end. We think the term is overloaded and prefer to use 
resultative for the latter group. 5 Halliday proposes two additional levels of sentence description ("metafunctions'), which operate in 
parallel to ideational structure: the interpersonal and the textual. For our present purposes, we can neglect them; for a broader scope of sentence generation, they are very important. 
406 
Stede Verb Alternations 
~11 poured water into the tank until it was filled. 
(xl / anterior-extremal 
:domain (x2 / directed-action :lex pour 
:actor (x3 / person :name jill) 
:actee (x4 / substance :lex water) 
:destination (x5 / three-d-location :lex tank)) 
:range (x6 / nondirected-action :lex fill 
:actee x5)) 
Figure 3 
SemSpec and a corresponding sentence. 
linguistic categories that directs the grammar in verbalizing objects (in the generator's 
input) in terms of these categories. Hence, the UM can be characterized as mirroring 
the distinctions made in surface linguistic realizations: Typically, any two distinct UM 
types correspond to some difference in English sentences. 
The largest part of the UM is devoted to processes, which are characterized by 
their verbalization patterns. For our purposes here, we need only a small fragment of 
the process hierarchy, namely the subtree of material processes. They can be charac- 
terized by the fact that English verbalizations of them in present tense typically use 
the progressive form, as in the house is collapsing (unmarked) as opposed to the house 
collapses (marked). They typically involve the participant roles "actor" and "actee "6 
but differ in terms of constraints on the types of the role fillers, and with respect to 
their realization in language. Material processes have two subgroups, one of which are 
nondirected-actions. They do not involve external agency and are mostly intransitive. 
With such processes, the actee is not a genuine participant, but rather an elaboration 
of the process. Verbs falling into this category are those of movement, of expressing 
skills, as well as support verbs like to take as in take a shower. The other subgroup, 
directed-actions, are always transitive, and they involve an external agent of the pro- 
cess. 
The upper model thus reflects the semantic distinctions made by the language, 
and the systemic-functional grammar takes care of the syntactic realization of these 
distinctions. Accordingly the lexicalization component we are proposing here is in 
charge of producing a sentence-semantic specification along the lines of ideational 
structure (using the upper model categories), such that the relevant decisions affecting 
sentence meaning can be controlled during lexical choice. As an example, Figure 3 
shows one of the SemSpecs and an English sentence that can be derived (as explained 
in Section 4) from the SitSpec given in Figure 2. Besides actor and actee, the role 
"destination" is used in the SemSpec; later we will also encounter "source." 
The UM is a good starting point, but in some respects the process classification is 
not quite fine-grained enough. A deficiency that is directly relevant for our treatment of 
alternations concerns the valency patterns of verbs, where some additional distinctions 
are needed. 
2.3.2 Valency. As introduced by Tesni~re (1959), valency refers to the distinction be- 
tween actants and circumstantials (central participants associated with the verb versus 
temporal, locational, and other circumstances). This separation is in principle widely 
accepted, but views differ on where to draw the line and how to motivate it. The 
notion of valency was further developed predominantly in German linguistics, with 
a culmination point being the valency dictionary of German verbs by Helbig and 
6 Actee is the upper model role that conflates what more often is called patient, theme, and goal. 
407 
Computational Linguistics Volume 24, Number 3 
Schenkel (1973). They made an additional distinction between obligatory and optional 
actants; Somers (1987, chapter 1) proceeded to propose six different levels of valency 
binding. He also pointed out that there are different opinions on the type of entities 
that are subject to a verb's valency requirements: some authors describe them by syn- 
tactic class, some by semantic deep cases, and some by their function (subject, object, 
etc.). 
Halliday (1985), in his classification, essentially adopts the basic Tesni~rian distinc- 
tion and suggests some semantic and syntactic criteria for deciding between actants, 
which he calls participants, and circumstances. Spatio-temporal information, for in- 
stance, is generally treated as a circumstance. As a syntactic indicator, for Halliday, 
participants are typically realized as nominal groups (with some obvious exceptions, 
as in say that x), and circumstances as prepositional phrases or as adverbs. But nei- 
ther this syntactic division corresponding to participants and circumstances (direct or 
indirect object versus adverbs or prepositional phrases), nor the semantic postulate 
that spatio-temporal aspects are circumstances, holds in general. Regarding spatial re- 
lationships, we find verbs that specifically require path expressions, which cannot be 
treated on a par with circumstances: Consider, for example, to put, which requires a 
direct object and a destination. Causative to pour requires a direct object as well as 
a path with either a source, or a destination, or both: pour the water from the can into 
the bucket. 7 Some verbs, as is well-known, can occur with either a path (Tom walked 
into the garden) or a place (Tom walked in the garden), and only in the garden can here be 
treated as a circumstance. And to disconnect requires a direct object (the entity that is 
disconnected) and a source (the entity that something is disconnected from), which 
can be omitted if it is obvious from the context: Disconnect the wire! 
As a step toward a more fine-grained distinction between participants and cir- 
cumstances, we adopt the three categories proposed by Helbig and Schenkel (1973) 
and thus distinguish between obligatory and optional participants on the one hand, 
and circumstances on the other. Moreover, we differentiate between requirements of 
process types (as encoded in the process taxonomy) and requirements of individual 
verbs, which are to be encoded in the lexical entries. In a nutshell valency (as a lexical 
property) supplements the participant/circumstance requirements that can be stated 
for types of processes. 
To encode the valency information, we introduce the partial semantic specification 
(PSemSpec) as one central component of lexical entries. The participant roles stated 
in the PSemSpec are either obligatory or optional; in the latter case they are marked 
with angle brackets: 
to disconnect 
PSS: (x / directed-action 
:actor A :actee B < :source C >) 
With obligatory participants, the verb is only applicable if the elements denoted by 
these participants are present in the input structure to be verbalized (the SitSpec). 
Optional participants need not be included in the verbalization: If they are present 
in the SitSpec, they may be omitted if there is some good reason (e.g., a stylistic 
preference); if they are not present in the SitSpec, the verb can be used anyway. The 
disconnect example illustrates that, in contrast to Halliday, we allow for verbs selecting 
path expressions, here as an optional complement. We can thus distinguish between 
7 Given a suitable context, though, the sentence She poured the wine is perfectly acceptable. But this usage seems to be restricted to a small class of digestible liquids. 
408 
Stede Verb Alternations 
cases like the following: 
• Tom disconnected the wire {from the plug}. To disconnect requires a source, 
but it can be omitted in a suitable specific context. 
• Sally ate. While to eat usually requires a direct object, it can also be used 
intransitively due to the strong semantic expectation it creates on the 
nature of the object--independent of the context. 
• Tom put the book on the table. To put requires a destination, and it cannot be 
omitted, no matter how specific the context. 
• The water drained from the tank {in the garage}. Locative circumstances like 
in the garage are not restricted to particular verbs and can occur in 
addition to paths required by the verb. 
The criterion of optionality, as indicated above, singles out the obligatory comple- 
ments from the other two categories. But how, exactly, can we motivate the distinction 
between optional participants and circumstances in our framework? By relating the 
PSemSpec to the SitSpec, via the denotation. In the disconnect case, for instance, the two 
items connector and connectee are both integral elements of the situation. The situation 
would not be well-formed with either of them absent, and the domain model encodes 
this restriction. Therefore, both elements also occur in the denotation of to disconnect, 
and a coindexed variable provides the link to the PSemSpec. Only when building the 
sentence SemSpec is it relevant to know that the connectee can be omitted. The con- 
nectee in the denotation therefore must have its counterpart in the PSemSpec--that is 
the source, but there it is marked as optional (see Figure 4 below). 
With circumstances, the situation is different: A SitSpec is complete and well- 
formed without the information on, for instance, the location of an event. Hence, a 
verb's denotation cannot contain that information, and it follows that it is not present 
in the PSemSpec, either. 
2.4 Lexical Entries 
We have introduced the two central components of lexical entries and now give a 
complete list of the components used in our system. The connotations are not directly 
relevant for our mechanism of handling verb alternations, therefore they will not be 
dealt with here. Salience assignment will be discussed in Section 4. 
Denotation: A partial SitSpec that defines the applicability condition of 
the lexeme: If its denotation subsumes some part of the input SitSpec, 
then (and only then) it is a candidate lexical option for the verbalization. 
Covering: The subset of the denotation nodes that are actually expressed 
by the lexeme. One of the constraints for sentence production is that 
every node be covered by some lexeme. 
Partial SemSpec (PSemSpec): The contribution that the lexeme can 
make to a sentence SemSpec. By means of shared variables, the partial 
SemSpec is linked to the denotation. 
Connotations: Stylistic features pertaining to formality, floridity, etc. 
Salience assignment (for verbs only): A specification of the different 
degrees of prominence that the verb assigns to the participants. 
409 
Computational Linguistics Volume 24, Number 3 
DISCONNECT 
CAUSER ~- :actor 
CONNECTOR ~- :actee 
CONNECTEE ~ <:source> 
POUR 
PATH-SOURCE 
OBJECT 
*PATH-DESTINATION 
*CAUSER 
--~ :actor 
> <:actee> 
substance-source 
durative-causative 
DRAIN 
OBJECT 
PATH-SOURCE 
*PATH-DESTINATION 
*CAUSER 
> :actor 
:~ <:SOUrCe> 
durative-causative 
locative/clear-intransitive 
resultative-causative 
MOVE/WALK 
OBJECT ~ :actor 
*PATH 
*CAUSER 
durafive-causative 
OBJECT -- 
*CAUSER 
OPEN 
> :actor 
resultative-causative 
SPRAY 
CAUSER ~ :actor 
OBJECT > :actee 
PATH-DESTINATION --->- :destination 
spray-load 
FILL 
CONTENT > :actor 
CONTAINER > :actee 
VALUE > <:destination 
(default)> 
*CAUSER 
stative-resultative 
resultative-causative 
LEAK 
PATH-SOURCE ~ :actor 
OBJECT <:actee> 
*PATI-I-DESTINATION 
substance-source 
Figure 4 
Excerpts from sample lexical entries for verbs. 
Alternation rules: (for verbs only): Pointers to lexical rules that represent 
alternations the verb can undergo (see Section 3). 
Morphosyntactic features: Standard features needed by the surface 
generator to produce correct utterances. 
Figure 4 gives excerpts from sample lexical entries, which demonstrate the linking 
between entities from the denotation and the PSemSpec. Notice that the linking is 
shown for the base form of the verb, which can be quite simple, as in open or move. 
Items appearing with an asterisk in front of them are optional in the SitSpec: for example, 
a SitSpec underlying an open-event is well-formed without a causer being present. 
These items get verbalized with the help of rules such as the alternation rules to which 
we turn in the next section. In the lexical entries, the names of applicable alternation 
rules are listed below the line. 
3. Alternations 
Having explained denotations and PSemSpecs, specifically for verbs, we can now turn 
to the task of accounting for the different alternations a verb can undergo. Under this 
heading, we will look both at the so-called transitivity alternations, which are charac- 
terized by a change in the number of participants (e.g., the causative), and at diatheses, 
410 
Stede Verb Alternations 
which only affect the mapping between the participants and syntactic realization (e.g., 
the passive). Thus, a variant such as topicalization does not qualify as an alternation, 
since the syntactic realization of the participants remains unchanged; they are merely 
reordered. The most comprehensive source of information on alternations is the com- 
pilation by Levin (1993); we will now look at some of the more prominent alternations 
listed there and characterize them in terms of changes in denotation and valency of 
the verbs. 
3.1 Alternations as Meaning Extensions 
A simple way of treating alternations is to use a separate lexical entry for every con- 
figuration, but that would clearly miss the linguistic generalizations. Instead, we wish 
to represent the common "kernel" of the different configurations only once, and use 
a set of lexical rules to derive the alternation possibilities. Jackendoff (1990) is con- 
cerned with this problem for a number of alternations; specifically, in his framework 
of lexical-conceptual structure (LCS) he seeks to explain the relationships between sta- 
five, inchoative, and causative readings of a verb. In Jackendoff's analysis, the forms 
are derived sequentially by embedding in the primitives INCH and CAUSE, respec- 
tively: 
• stative: BE(\[Thi,x \](A), \[INa \[Thing \]A \]) 
• inchoative: INCH \[BE(\[Thing \](A), \[INd \[Thing \]A \])\] 
• causative: CAUSE(\[Thi,g \]A, INCH \[BE(\[Thing \](A), \[INd \[Thing \]A \])\] 
For our NLG purposes, the idea of deriving complex verb configurations from 
more basic ones is attractive, but it is necessary that we relate verb meaning to our 
explicit treatment of event structure, instead of masking that structure with a primitive 
such as INCH. When verbalizing a SitSpec, we first have to determine candidate 
lexemes, i.e., match the SitSpec against lexicon entries; having only one lexical entry 
for a verb reduces the search space considerably. Moreover, since the verb entry will 
be the most basic form, its denotation is relatively simple and therefore the matching is 
inexpensive. Finding more complex verb configurations will then require some further 
matching, but only locally and to those verbs that have already been determined as 
verbalization options. 
In general the idea is to see verb alternations not just as relations between different 
verb forms, but to add directionality to the concept of alternation and treat them as 
functions that map one into another. From this viewpoint, there are two groups of 
alternations: (1) Alternations that do not affect the denotation of the verb. Examples 
are the passive or the substance-source alternation (The tank leaked oil; Oil leaked from 
the tank): The truth conditions do not change. (2) Alternations that do change the 
denotation of the verb. 
The second group is the critical one, because if we derive verb configurations from 
others and rewrite the denotation in this process, it has to be ensured that the process 
is monotonic. 8 Therefore we define directionality for group (2) to the effect that an 
alternation always adds meaning: The newly derived form communicates more than 
the old form the denotation gets extended. This notion is different from the standard, 
nondirectional way in which alternations are seen in linguistics; to label the difference, 
we call alternations of group (2) extensions. In this section, we will introduce a number 
8 Monotonicity might be too strong a requirement for alternations in general, though. See Section 5.3. 
411 
Computational Linguistics Volume 24, Number 3 
of extension rules for which we can give a clear definition in terms of Aktionsart 
features, as they were introduced in Section 2.2. These rules extend the denotation 
of a verb and rewrite its PSemSpec in parallel to reflect the change in valency; the 
result is a new verbalization option, which can differ from the previous one in terms 
of coverage or attribution of salience (see Section 4). The rules will be conveniently 
simple to state, thanks to the upper model, which provides the right level of abstraction 
from syntax. 
To illustrate the goal we return to the example of Tom removing the oil from an 
engine. If a SitSpec encodes this situation, then to drain is a candidate lexeme. While 
it can appear in a number of different configurations, we wish to match only one of 
its forms against the SitSpec, though. This is the most basic one, denoting an activity: 
The oil drained from the engine. Here, the case frame of the verb has to encode that 
from the engine is an optional constituent. Now, an extension rule has to systematically 
derive the causative form: Tom drained the oil from the engine. And also from the first 
configuration, another rule derives the resultative reading, which adds the information 
that the engine ended up empty: The engine drained of the oil. Here, of the oil is an optional 
constituent. To this last form, a causative extension can apply and yield Tom drained 
the engine of the oil. 
To compute these configurations automatically, such that valency and meaning are 
changed in parallel, we define an alternation or extension rule as a 5-tuple with the 
following components: 
NAM: a unique name; 
DXT: extension of denotation; 
C0V: additions to the covering-list; 
ROC: role changes in PSemSpec; 
NR0: additional PSemSpec roles and fillers. 
The DXT contains the denotation subgraph that the new verbalization has in addition 
to the old one. The syntax is, of course, the same as that of the denotation of a lexical 
entry. Specifically, it can contain variables; these can co-occur in the C0V list: the items 
that the new verbalization covers appear in addition to those of the old one. R0C is a 
list of pairs that exchange participant role names or the UM type in the PSemSpec; this 
replacement can also change optionality. For example, (< :actee > :actor) means 
"replace the term :actee in the PSemSpec of the old verbalization, where it was 
optional, with :actor, which is not optional." Finally, NR0 contains new roles and 
fillers that are to be added to the new PSemSpec; these will also contain variables 
from the denotation extension. 
Applying such a rule to a verbalization option vo works as follows: Add the 
contents of DXT to the denotation of vo, and match the new part against the SitSpec. If 
it matches, make a copy vo t of vo and assign it a new name as well as the denotation 
just formed. Add the C0V list, which has been instantiated by the matching, to the 
covering-list of vo'. Exchange the role names in the PSemSpec of vo ~ as prescribed by 
ROC, and, importantly, in the order they appear there. Finally, add NR0 to the PSemSpec. 
In the following, we first give an example for a rule that changes only the PSem- 
Spec without affecting the denotation. Afterwards, we describe those alternations that 
change the Aktionsart of the verb and thus the form of the SitSpec expressed (as 
discussed in Section 2.2): the stative-resultative, causative, and locative alternations. 
412 
Stede Verb Alternations 
Before introducing these rules, it should be emphasized that we do not provide appli- 
cability conditions for the alternation and extension rules, which would inspect some 
verb denotation and on that basis decide whether an alternation can apply; instead, 
the rules are triggered directly from the lexical entry of a verb. Whether general ap- 
plicability conditions can be specified, so that the rules need not be attached to each 
individual verb, is a central open research question that linguistic alternation research 
is concerned with. 
3.2 Encoding Alternation Rules 
The best-known alternation that affects only the valency of the verb is the passive, 
which we do not investigate here. Instead, we show one alternation that is particularly 
relevant for verbs in the domain of substances and containers. 
Substance-Source Alternation. Example: The tank leaked water / Water leaked from the tank. 
This is an alternation discussed by Levin (1993); to make use of it here, we have to 
add directionality and declare one of the two configurations as more basic. Levin lists 
verbs of "substance emission" as undergoing it, for example drip, radiate, sweat, and 
leak. 9 To decide on the more basic form, we use the fact that in The tank leaked water 
the water is an optional constituent, and hence the minimal configuration of the verb 
is The tank leaked. With the from configuration, no deletion is possible. 
As a representative of the verb class, we show the denotation and PSemSpec of to 
leak: 
DEN: (leak (OBJECT A) 
(PATH (SOURCE B))) 
PSS: (x / nondirected-action 
:lex leak :actor B < :actee A >) 
The following alternation rule applies to all these substance emission verbs and 
derives the from configuration: 
NAM : substance-source 
DXT: () 
C0V: () 
R0C: ((:actor :source) 
(< :actee > :actor)) 
NRO: () 
Let us now consider several alternations that change denotation, and hence are exten- 
sions. 
Stative-Resultative. Example: Water filled the tank / The tank filled with water. In discussing 
verbs that denote a state, Jackendoff (1990) points out that fill, cover, surround, and 
saturate can describe either a state or an inchoative event, and encodes the difference 
with the primitive INCH we have shown in the introduction to this section. Our goal 
is to do without the primitive, and to define the change in terms of the Aktionsart of 
the verb; to this end, we use resultative in place of inchoative (see Section 2.2). 
9 Unnoticed by Levin, to leak can also be a verb of substance "intrusion," as in The camera leaked light. This 
reading, which we do not handle here, reverses the directionality of the path involved. 
413 
Computational Linguistics Volume 24, Number 3 
On a similar matter, Levin (1993) describes the "locatum subject" alternation; 
which for instance holds between I filled the pail with water and Water filled the pail. 
It thus relates a causative and a noncausative form. Levin states that the alternation 
applies to a class of "fill verbs," of which there are many more than the four given by 
Jackendoff. Her alternation is not exactly the one we need here, since it also involves 
a causative form; deriving the causative is a separate step in our framework. 
What we need here is a mixture of Jackendoff's and Levin's insights: Several of 
Levin's fill verbs can be both transitive and intransitive, and some of the intransitive 
readings denote 'to become Xed'. Among these verbs are fill, flood, soak, encrust, or 
saturate: The kitchen flooded with water means the same as The kitchen became flooded with 
water. For this subgroup of the fill verbs, we define an extension rule that derives a 
resultative reading from a state reading. 1° Notice that this is different from Levin's 
locatum subject alternation, since it does not involve a causer. 
NAM : stative-resultative 
DXT: (event (Y (ACTIVITY X))) 
C0V: (X Y) 
R0C: ((:actor :inclusive) 
(:actee :actor) 
(directed-action nondirected-action) 
NRO: () 
To illustrate the rule with an example, consider the denotation and PSemSpec of 
the state reading off ill: 
DEN: (fill-state (CONTAINER A) 
(C0NTENT B) 
(VALUE C) ) 
PSS: (x / directed-action :lex fill 
:actor B :actee A < :destination C >) 
When matching it against a SitSpec with a tank and water, this yields the verbalization 
The water filled the tank, covering only the post-state of the SitSpec. Now, the alternation 
rule extends the denotation to also cover the event and the activity that brings the 
filling about. Applying the changes to the PSemSpec results in 
(X / nondirected-action :lex fill 
:inclusive B :actor A < :destination C >) 
which corresponds to the sentence The tank filled with the water. 
A few stative verbs cannot be resultative without being also causative. Consider 
to cover in these examples from Jackendoff: 
Snow covered the ground. 
,The ground covered with snow. 
Bill covered the ground with snow. 
10 The two roles "inclusive" and "of-matter" (used later) are the roles used by Penman to realize the 
desired structure, but they are not very good descriptions of these semantic relationships. For a more 
systematic treatment, for instance along the lines of Somers (1987), the upper model needs to be 
extended. See Section 4.1. 
414 
Stede Verb Alternations 
For these, a stative-culmination extension derives the resultative + causative form di- 
rectly from the stative one. The rule is similar to the one given above, so we do not 
show it here. 
Causative Extensions. Example: The napkin soaked ! Tom soaked the napkin. Levin discusses 
a causative/inchoative alternation that applies to a large number of verbs. The class 
formed by them is somewhat heterogeneous with respect to Aktionsart, though; it 
contains, for example, to turn as well as to open. The former is in its basic form durative 
(The wheels turned), and the latter transformative (The door opened). Accordingly, we split 
the alternation in two, which only differ in the DXT component, reflecting the difference 
in Aktionsart. The durative-causative extension adds a causer to the denotation and 
makes the former :actor the new :actee. It equally applies to semelfactive verbs 
denoting a momentaneous activity: The bell rang / The visitor rang the bell. The resultative- 
causative extension also covers the activity, because Tom opened the door expresses that 
Tom did something to achieve the change of state. The causer itself is not covered 
though, because it still has to be verbalized separately. 
NAM : durative-causative 
DXT: (activity (CAUSER X)) 
COY: () 
ROC: ((:actor :actee)) 
NRO: (:actor X) 
NAM: resultative-causative 
DXT: (event (ACTIVITY (X (CAUSER Y)))) 
COY: () 
ROC: ((:actor :actee)) 
NRO: (:actor Y) 
The first rule derive~for examplG Tom walked the dog from The dog walke~ andthe 
second Tom closed thedoor ~omThe doorclosed. 
Locative Extensions. Example: (a) Sally sprayed the wall with paint. / (b ) Sally sprayed paint 
onto the wall. The locative alternation has been studied by lexical-semanticists exten- 
sively. Its characteristic is that configuration (a) of the verb conveys that something is 
performed in a "complete" or "holistic" manner, whereas configuration (b) lacks this 
facet of meaning. Levin points out that this alternation has received much attention 
in linguistics research and notes that, in spite of the efforts, a satisfactory definition of 
the holistic facet has not been found. Jackendoff, in his treatment of the alternation, 
suggests encoding the holistic feature in a primitive: The function ONd is a deriva- 
tive of ON and means that something "distributively" covers a surface, e.g., the paint 
covers all of the wall. Introducing a primitive, though, amounts to conceding that 
no explanation in terms that are already known can be given. We cannot solve the 
question of "holisticness," either, but we want to point to the fact that the two verb 
configurations correlate with a change in Aktionsart: Sally sprayed paint onto the wall is 
durative (she can do it for two hours), whereas Sally sprayed the wall with paint is trans- 
formative (she can do it in two hours). That observation leads us to propose that the 
example is best analyzed as involving a mere activity in the with configuration, and an 
additional transition in the onto configuration. Support for this analysis comes from 
Pinker (1989), who postulates a change in meaning when moving from one configura- 
tion to the other: In (b) above, Sally causes the paint to move onto the wall whereas 
in (a), Sally causes the wall to change its state by means of moving the paint onto it. 
415 
Computational Linguistics Volume 24, Number 3 
Sally sprayed paint onto the wall. 
(spray-i (CAUSER sally-l) 
(OBJECT paint-l) 
(PATH (path-i (DESTINATION wall-l)))) 
Sally sprayed the wall with paint. 
(event-I (PRE-STATE (covered-state-I (OBJECT wall-l) 
(VALUE (not 'covered)))) 
(ACTIVITY (spray-i (CAUSER sally-l) 
(OBJECT paint-l) 
(PATH (path-i (DESTINATION wall-l))))) 
(POST-STATE (covered-state-I (OBJECT wall-l) 
(VALUE 'covered)))) 
Figure 5 
SitSpecsfor configura~ons of ~ spray. 
Pinker sees (a) as derived from (b) and suggests, as a constraint on the applicability of 
the alternation, that the motion (here: spray) causes an effect on the surface/container. 
While we decided not to discuss applicability conditions here, we support the idea 
that the difference between (a) and (b) can be expressed with an additional change of 
state. In our framework, we thus assign two different SitSpecs to the sentences, one 
activity and one event, as shown in Figure 5. 
The crucial point now is that the first SitSpec is fully embedded in the second; 
this is in correspondence with the truth conditions: If Sally has sprayed the wall with 
paint, then she also has sprayed paint onto the wall. To generalize the correspon- 
dence to an extension rule, we need to assume in the domain model a concept like 
completion-state, which is to subsume all those states in the domain model that have 
"extreme" values: an empty bucket, a fully loaded truck a completely covered sur- 
face, and so forth. The exact interpretation of completion-state is the open question 
that Levin (1993) referred to, and that Jackendoff treated with his d subscript. We do 
think, though, that an abstract state in the domain model, which subsumes a range 
of the concrete states, is preferable to introducing a primitive on the linguistic level 
(unless the primitive is relevant for other linguistic phenomena as well). 
The following alternation rule applies to durative verb readings that denote ac- 
tivities of something being moved to somewhere, and extends them to also cover 
the post-state, which must be subsumed by completion-state. In this wan it derives 
reading (a) from (b) in the spray example, and analogously for the other verbs un- 
dergoing the alternation, e.g.: Tom loaded hay onto the wagon ! Torn loaded the wagon with 
hay; Jill stuffed the feathers into the cushion ! Jill stuffed the cushion with the feathers. The 
PSemSpec is modified as follows: The former :destination (wall) becomes the new 
:actee, whereas the former :actee (paint) now fills the role < :inclusive >, and is 
optional there, because Jill stuffed the cushion is also well formed. 
NAM : locative-transitive 
DXT : (event 
(MOVE (OBJECT X) 
(PATH (DESTINATION Y))) 
(POST-STATE (Z completion-state (OB3ECT Y)))) 
COY : (Z) 
ROC: ((:actee < :inclusive >) 
(:destination :actee)) 
NRO: () 
416 
Stede Verb Alternations 
Levin distinguishes two kinds of locative alternation: the spray/load alternation 
just discussed and the clear (transitive) alternation. The latter applies only to the verbs 
clear, clean, drain, empty and can be seen as the "semantic inverse" of the spray/load 
alternation, because one group of verbs denotes activities of placing something some- 
where, and the other describes activities of removing something from somewhere; but 
both have the same holistic effect in one of the verb configurations. Thus, the rule for 
the clear-alternation is very similar to the one just shown. It derives, for example, Tom 
drained the container of the water from Tom drained the water from the container, n 
NAM: clear-transitive 
DXT : (event 
(MOVE (OBJECT X) 
(PATH (SOURCE Y))) 
(POST-STATE (Z completion-state (OBJECT Y)))) 
C0V: (Z) 
ROC: ((:actee < :of-matter >) 
(:source :actee)) 
NR0: () 
The clear verbs, except for to clean, can in addition be intransitive, and Levin states 
a separate alternation for them. For to drain, the first configuration is The water drained 
from the tank, and the second is either The tank drained or ?The tank drained of the water. 
According to Levin, "the intransitive form may be best in the absence of the of-phrase" 
(Levin 1993, 55). The SitSpec denoted by the first configuration is: 
The water drained from the tank. 
(move-i (OBJECT water-l) 
(PATH (path-I (SOURCE tank-l)))) 
Note that our durative-causative extension rule given above applies in this case and 
extends the coverage of the SitSpec to one corresponding to Tom drained the water from 
the tank. A rule that is parallel to that for the transitive case is given below; it derives 
: of-matter > is optional, we can also produce ?ThetankdrainedofthewateG sincethe < 
the preferred The ~nk drained. 
NAM: locative/clear-intransitive 
DXT: (event 
(MOVE (OBJECT X) 
(PATH (SOURCE Y))) 
(POST-STATE (Z completion-state (OBJECT Y)))) 
C0V: (z) 
ROC: ((:actor < :of-matter >) 
(:source :actor)) 
NRO: () 
3.3 Deriving Alternations Successively 
The extension rules, as we have introduced them above, constitute a framework for 
systematically deriving more complex verb configurations from simpler ones; the out- 
put produced by one rule serves as input to another. Figure 6 provides a synopsis: The 
11 We ignore the role of the definite determiner here, which in fact has critical influence on the holistic 
interpretation of mass nouns. See, for example, White (1994). 
417 
Computational Linguistics Volume 24, Number 3 
• "-(?-tat'~ x) l (~cti' 
(event(PRE-STATEX)--~ r 7 
.......... !~t?g:.s.7:~?!:x!!.. t / 
(event(ACTWITY X) I 
..... .._(.P.O..S.T:S.T..A.~.X)) .......... \] RESULTATIVE 
J 
c uSl~t'l~aefi°n~ resultative- 
(event(PRE-STATE X) \[ } (ACTIVITY(CAUSER Y)) \] 
\[ RESULTATIV E+CAUSATIVE \] 
~y.X)..\] ,TIVE \] 
durative- causative 
DURATIVE+CAUSATIVE \] 
causative spray/~ 
Figure 6 
Dependency of extension rules. 
boxes contain the denotation patterns that correspond to the Aktionsart feature, and 
the rules transform a configuration with one Aktionsart into another. In this graph, 
every verb base form has an entry point corresponding to the Aktionsart of its most 
basic configuration. Examples: to fill is stative, to drain is durative, to open is transfor- 
mative, to remove is resultative + causative. The "double box" in the middle is the entry 
point for both transformative and resultative verbs, but the incoming arrows produce 
resultative forms. From the entry point of a verb, arcs can be followed and rules ap- 
plied if the respective alternation is specified in the lexical entry. The six categories 
account for the Aktionsart features listed in Section 2.2, and the rules take care of 
possible shifts between them. Thus, the full range of SitSpecs that our ontology allows 
is being covered. 
To illustrate the functionality, we return to the example of to drain. Figure 7 shows 
how the extension rules successively derive the various configurations. Apart from the 
passive, this is the complete "alternation space" of to drain according to Levin's (1993) 
catalogue. Notice that the examples given also cover the four different drain clauses 
needed to produce the alternative sentences given in (1) in the introduction. 
4. Implementation: Two-Step Sentence Generation with MOOSE 
The MOOSE sentence generator grew out of experiences with building the TECHDOC 
system (RSsner and Stede 1994), which produces instructional text in multiple lan- 
guages from a common representation. Specifically, MOOSE accounts for the fact that 
events can receive different verbalizations even in closely related languages such as 
English and German. It is designed as a sentence generation module that pays at- 
tention to language-specific lexical idiosyncrasies, and that can be incorporated into 
418 
J 
Stede Verb Alternations 
Denotation: (activity (OBJECT A) 
(PATH (SOURCE B))) 
PSemSpec: (xl / nondirected-action :lex drain 
:actor A :source B) 
(0) The oil drainedfromthe engine. 
Locative/clear-intransitive of (0): 
Denotation: (event (ACTIVITY (OBJECT A) 
(PATH (SOURCE B))) 
(POST-STATE (C (OBJECT B)))) 
PSemSpec: (xl / nondirected-action :lex drain 
:of-matter A :actor B) 
(1) The engine drained ofthe oil. 
Durative-causative of (0): 
Denotation: (activity (OBJECT A) 
(PATH (SOURCE B)) 
(CAUSER C)) 
PSemSpec: (xl / directed-action :lex drain 
:actee A :source B :actor C) 
(2) Tom drained the oil from the engine. 
Resultative-causative of (1): 
Denotation: (event (ACTIVITY (OBJECT A) 
(PATH (SOURCE B)) 
(CAUSER C) ) 
(POST-STATE (C (OBJECT B)))) 
PSemSpec: (xl / directed-action :lex drain 
:of-matter A :actee B :actor C) 
(3) Tom drained the engine of the oil. 
Figure 7 
Derivation of drain configurations. 
a larger-scale text generator} 2 In the following, we first describe the overall system 
architecture, then discuss the process of lexicalization in some detail, and finally turn 
to the selection of a verb alternation on the basis of salience parameters. 
4.1 System Architecture 
Figure 8 provides an overview of the architecture of MOOSE. The generator assumes a 
language-neutral level of event representation, the situation specification SitSpec (see 
the example in Figure 2). The SitSpec instantiates concepts from a domain model, 
which is implemented in the KL-ONE language LOOM (MacGregor and Bates 1987). 
Using the denotations of the lexicon entries of the target language, the lexical options 
for verbalizing the SitSpec are determined. In the next step, for verbs, the applicable 
alternations and extensions are computed and added to the set of options. Then a 
language-specific semantic specification SemSpec (see the example in Figure 3) is con- 
structed in accordance with generation parameters pertaining to brevity, salience, and 
stylistic features. The SemSpec is then handed over to a surface generator: Penman 
(Penman group 1989) for English, and a variant developed at FAW Ulm for German. 
The SemSpec language is a subset of the input representation language that was 
developed for Penman, the sentence plan language (SPL) (Kasper 1989). An SPL ex- 
pression consists of variables, types, and case roles; an example was given in Figure 3. 
Penman and SPL are based on the upper model (UM) (Bateman et al. 1990) introduced 
12 Fr6hlich and van de Riet (1997) describe how MOOSE is employed in the generation component of an 
information system. 
419 
Computational Linguistics Volume 24, Number 3 
Domain 
Model 
°o& '-.. ..:, . . 
i .•/ i~: 
SitSpec 
Lexicon 
Morphosyntax Morphosyntax Mowhosyntax 
Alternations Alternations Alternations 
Partial SemSpec Partial SemSpec Partial SemSpec 
Connotation Connotation Connotation 
Denotation Denotation Denotation 
Alternation/ 
Extension 
rules -- Ve lzatlon@ ................ L 
Local yor eredl 
verbalization //~N Upper Model 
options F_JG 
(4) I Unijicati°n of ~ ............................. 
Well-formed, .' 
complete, ..' 
preferred / 
SemSpec 
f Su,~a--~enera"on 1 
(5) L Penman EZG J /\, 
English German 
sentence sentence 
Figure 8 
MOOSE system architecture. 
in Section 2.3.1. For any type appearing in an SPL, Penman needs to know by which 
UM type it is subsumed, so that appropriate generation decisions can be made. 
The way we use Penman and the UM in the Moose architecture is somewhat 
different from the original Penman conception. In Penman, the domain model was 
supposed to be subsumed by the UM, which indeed simplifies generation from input 
that uses domain model concepts. However, the range of alternative verbalizations 
420 
Stede Verb Alternations 
that can be produce d from the same input is seriously limited under this approach 
(see Stede and Grote \[1995\]), and therefore Moose opts for a complete separation of 
DM and UM; they are distinct taxonomies. Consequently, as opposed to a general SPL 
term used in Penman, a SemSpec used in Moose must contain only upper model 
concepts and no domain model concepts. 
Furthermore, since our system takes lexicalization as the decisive task in mapping 
a SitSpec to a SemSpec, the UM concepts referred to in a SemSpec must be annotated 
with :lex expressions. Thus, a SemSpec is a lexicalized structure, and accordingly, 
Moose interprets the upper model as a taxonomy of lexical classes. This contradicts the 
Penman philosophy of viewing the UM as abstract semantics and clearly distinct from 
the generation grammar, which in accordance with systemic-functional linguistics is an 
integrated lexicogrammar, with "lexis as most delicate grammar" (Hasan 1987). This 
idea, however, has been a theoretical rather than a practical one, and lexical matters 
thus have not been a strong point of Penman. For instance, the distinction between 
obligatory and optional participants of a verb was quite blurred. Also, Penman allowed 
only for very simple lexical choice mechanisms, as it assumed a straightforward one- 
to-one mapping between concepts and words. MOOSE overcomes these problems by 
assigning a central role to the lexicon, placing a lot of information in it, and taking it as 
the crucial device for the SitSpec-SemSpec mapping. SemSpec, then, is an intermediate 
level of representation that reflects sentence semantics and that mediates between 
the language-neutral conceptual representation and linguistic realization. The simple 
form of our alternation rules shown in the last section, which abstract over syntactic 
realization, demonstrates the utility of SemSpec as a level of description. 
In practice, our aim to upgrade SPL from a convenient input notation of a front- 
end NLG module to a systematic and well-motivated level of description involves not 
only building Moose "around" Penman, but also making some changes to the upper 
model and the generation grammar. But for the purposes of this paper, which focuses 
on the semantics of verb alternations and their role in NLG, we avoid dealing with 
Penman's internals and rather treat it as a "black box." 
4.2 Lexicalization 
In order for serious lexical choices to be possible, the first step of lexicalization in 
Moose consists of determining the set of verbalization options: all the lexemes whose 
denotations can potentially cover some part of the input SitSpec. Since we represent the 
internal structure of events, the denotation of a lexeme need not be a single concept; 
instead, it can be a complete configuration of concepts and roles. The consequences 
are a higher computational cost in finding lexical options, but also a greater flexibility 
in finding different verbalizations of the same event. As an example, consider the 
denotation of the causative reading of tofilh 
(event (PRE-STATE (fill-state (VALUE (not 'full)) 
(CONTAINER A))) 
(ACTIVITY (CAUSER B)) 
(P0ST-STATE (fill-state (VALUE < D 'full >) 
(CONTAINER A) 
(C0NTENT C) ) ) ) 
Given some input SitSpec involving filling, the variables of the denotation are bound 
to instances or atomic values of the SitSpec when it is matched against the denotation. 
The filler of the value role in the post-state appears in angle brackets because it is a 
default value. The accompanying partial SemSpec of to fill contains the same variables: 
421 
Computational Linguistics Volume 24, Number 3 
(x / directed-action :lex fill 
:actor B :actee A :inclusive C <:destination D>) 
When the denotation is matched against a SitSpec, the variable bindings are prop- 
agated to the partial SemSpec; and when it is later unified with the partial SemSpecs 
corresponding to the other elements, a complete SemSpec results, from which Penman 
produces a sentence like Jill filled the tank with oil. (If the value is different from 'full, 
it also gets verbalized, such as in Jill fflled the tank to the second mark.) 
Importantly, the matching between denotations and SitSpec does not test for iden- 
tity, but for subsumption--it exploits the functionality provided by LOOM. In this 
way, the selectional restrictions of verbs are checked when the lexical options are de- 
termined. Moreover, the matcher finds not only the most specific word, but also the 
applicable more general candidates, which is helpful, for instance, in achieving stylistic 
effects, and in avoiding undue repetitions of the same specific term. 
Since we are using relatively fine-grained representations for SitSpecs and denota- 
tions, the generation of variants in incorporation is enabled by the covering mechanism 
in conjunction with the subsumption check. In the example go by plane/fly, the general 
verb to go covers only the move concept, and the role instrument-plane is left to be 
expressed by a prepositional phrase; whereas the specific verb toffy covers the whole 
configuration. In this fashion, quite different coverings of the input SitSpec are possi- 
ble; for instance, MOOSE produces Tom poured water into the tank until it was full and Tom 
filled the tank with water (amongst others) as paraphrases of the same event. 
After the initial matching between denotations and SitSpec, the various alterna- 
tions are computed for those verbs whose base form has been found as a candidate 
(step 2 in Figure 8). Their lexical entry specifies which alternation/extension rules 
apply, and they are executed sequentially, as outlined in the previous section, and 
demonstrated for to drain in Figure 7. For any extension rule that adds new items to 
the denotation, the new material is matched against the SitSpec to ensure that the 
alternation is applicable, and to compute the additional covering. In this way, all the 
applicable alternated forms of a verb are added to the pool of verbalization options. 
The set of all lexemes that successfully matched some part of the SitSpec, together 
with the alternated verb forms, constitute the search space for constructing an appro- 
priate SemSpec. The options are first brought into an order of preference (step 3 in 
Figure 8) according to various parameters such as the desired salience assignment, 
which is explained in the next section. Considering the options in this order, a com- 
plete and well-formed SemSpec is built from the partial SemSpecs that are associated 
with some of the lexical options--those that collectively cover the entire SitSpec and 
thus will take part in the sentence. This is done by a unification process driven by the 
candidate verb options; recall that their PSemSpec consists of an upper model process 
and the mappings from situation elements to process participants, which is achieved 
by co-indexing with positions in the denotation. By means of sharing this information 
between denotation and PSemSpec, the lexicon entries serve as a "bridge" between 
the SitSpec to be verbalized and the intermediate representation SemSpec. 
For more details on the kinds of mono- and multilingual variation produced by 
MOOSE, and on the lexicalization algorithm, see Stede (1996b). 
4.3 Producing Salience Variation with Alternations 
Having explained the basic machinery of MOOSE, we now demonstrate how the gen- 
erator can make an informed choice among the set of possible verb alternations on 
the basis of a salience parameter, Since a full-fledged treatment of the role of salience 
422 
Stede Verb Alternations 
is well beyond the scope of this paper, 13 we will merely sketch a possible division of 
labor between text planning and sentence planning, and then describe the role of verb 
alternations as one means of realizing generation goals related to salience. Specifically, 
we show under what circumstances the various drain sentences given in the examples 
in (1) in Section 1 (and in Figure 7) are produced by MoosE. 
When text planning has been completed and linearization as well as the "chunk- 
ing" of the material into sentence-size pieces is accomplished, Moose takes over to 
perform the necessary sentence planning, which includes lexicalization. Thus, text 
planning has produced a sequence of SitSpecs, which may be enriched with infor- 
mation pertaining to the relative salience of the elements. This information can result 
from constraints on theme development, from rhetorical strategies, or from other con- 
siderations at the discourse level. For instance, in the sample texts given in (1) at the 
beginning of Section 1, different theme developments are responsible for the different 
usages of to drain. 
For sentence planning, we assume that individual nodes of a SitSpec can have a 
foreground, background, or optional label attached to them (but they need not). Then, 
a realization is to be found that signals the differences in prominence on the linguistic 
surface. In general, there is no one-to-one correspondence between the configuration 
of salience labels and linguistic realization, though. Instead, we view salience goals as 
goals that the generator tries to fulfill if possible, similar to certain stylistic goals (see 
Stede \[1996a\]). Thus, generation becomes a matter of constraints (say the right thing) 
and preferences (try to say it in a particular way), similar to Hovy's (1988) distinction 
between "prescriptive" and "restrictive" planning. 
What, then, is the role of verb alternations in assigning different degrees of salience? 
Talmy (1988) listed a number of morphological and syntactic means to distribute 
salience across the elements of a clause. For instance, he suggested the hierarchy sub- 
ject > direct object > indirect object > oblique, ranging from the most salient to the 
least salient. From a slightly different perspective, Kunze (1991) was concerned with 
differences in salience between similar verbs. He advanced the view that they share a 
common underlying base form and differ, inter alia, in distributing salience via their 
case roles. For our purposes here, we can adapt these insights (with some simplifica- 
tion) and state that an element is placed in the foreground if it is mapped to the role 
actor (best) or to actee (second best). Correspondingly, it is placed in the background 
if it corresponds to a circumstance, i.e., a role that is not part of the verb's case frame. 
Now, consider again the sentences in Figure 7. On the one hand, (0) and (1) omit 
the fact of Tom's causing the event, and hence are preferred only if the respective 
SitSpec node is labeled as optional. On the other hand, (0, 2) and (1, 3) differ in that 
the former render the oil prominent, while the latter emphasize the engine. Figure 9 
shows the common SitSpec underlying the four sentences, and a set of salience labels 
attached to three nodes, where the numbers correspond to the target sentences. For 
example, when sentence (1) is the preferred output, the SitSpec would have an opt 
label at node tom-1 and an fg label at node engine-1. 
For any verbalization option, base forms and alternations alike, the number of 
fulfilled salience goals can be computed straightforwardly: Since variables in denota- 
tions and PSemSpecs are co-indexed, we can determine for every salience label in the 
SitSpec how the corresponding element participates in the SemSpec. Using the criteria 
given above, preference values result for the various options, and they are factored 
into the overall preference ranking of the verbalization options. All other things being 
13 Pattabhiraman (1992) devoted a dissertation to the topic of salience in NLG. 
423 
Computational Linguistics Volume 24, Number 3 
> ' full 
/ ~ > oil-I 
event- 1 drain- 1 ~/ 
engine-1 
(l,3):fg fill-state-2 
~ " > 'empty 
Figure 9 
SitSpec representing a drain event. 
equal, the verb alternation that accomplishes the best realization of the salience labels 
outranks the other options and thus gets selected for building the SemSpec. Again, 
notice that other syntactic and morphological means (e.g., expressing an element with 
a separate word versus incorporating it into another word) for assigning salience can 
be integrated into this scheme. 
5. Summary and Related Work 
We have proposed an NLG framework, together with suitable representation schemes, 
that can systematically produce a range of verb alternations from a common underly- 
ing input representation and select the most appropriate form on the basis of salience 
parameters. Productive rules derive the more complex forms from a basic one, which 
is the only one that needs to be stated in the lexical entry of the verb. We have focused 
on those alternations that affect the Aktionsart of the verb: They imply a type change 
in aspectual classifications such as those of Bach (1986). 
For generation, our approach uses two distinct ontologies: a language-neutral do- 
main model for event categorization, and a language-specific taxonomy, the upper 
model developed by Bateman et al. (1991) on the basis of Halliday's (1985) work. The 
lexicon acts as the mediator between these two realms and serves to map a concep- 
tual input representation to a semantic sentence specification, which can be further 
processed by a front-end realization component. Within this framework, multilingual 
generation is possible once language-specific upper models and front-ends are used 
(but multilinguality was not addressed in this paper). The approach has been imple- 
mented in the MOOSE system, which uses the Penman generator (Penman Group 1989); 
MOOSE can serve as a plug-in sentence production module to a larger text generator. 
The examples discussed in this paper (including the alternated forms) were gener- 
ated from a domain model that encodes knowledge about automobile engines, tanks, 
and related liquids. It consists of 150 concepts and relations; the associated English 
lexicon has 200 words, including 50 verbs. Given the nature of the work, which de- 
pends on quite fine-grained representations in both domain model and lexicon, it is 
difficult to make statements on how well the approach "scales up." Large-scale (au- 
tomatic) acquisition of dictionary entries typically does not result in representations 
of the kind needed here, and furthermore, the domain model needs to be developed 
in tandem with the lexicon. The precise shape of such models, on the other hand, 
also depends on the specific application the generator is used for; even though some 
steps towards standardization in ontologies are being taken, this is still a bottleneck 
in knowledge-based NLG. 
424 
Stede Verb Alternations 
In the following, we compare our approach to some related work on verb alter- 
nations and on lexicalization in NLG. Finally, we draw some conclusions as to the 
overall scope of the work and its utility for NLG. 
5.1 Alternations 
Starting from the aspectual categories proposed by Bach (1986), the verb classifications 
of Levin (1993), and the lexical representations given by Jackendoff (1990), we have 
developed a new synthesized approach for dealing with verb alternations that affect 
the Aktionsart of verbs. Our ontology for input representations and the specifications 
for lexical meaning have benefited from the earlier work just mentioned but essentially 
constitute a new framework in which the specific alternation/extension rules could be 
formulated. 
The utility of these rules demonstrates the importance of defining a place for fine- 
grained lexical-semanfic representations in language generation. To our knowledge, 
no other generator can systematically derive the various forms for the alternations 
discussed in this paper. Recentl.~ Dorr and Olsen (1996) suggested using verb rep- 
resentations based on Jackendoff's (1990) LCSs for NLG; specific kinds of LCSs are 
proposed to represent different classes of verbs on the basis of telicity. Rules are pro- 
posed that relate telic and atelic versions of the same verb. The central difference of 
our approach is our distinction between SitSpec and SemSpec, and thus between de- 
notation and PSemSpec in the lexical entries. Dorr and Olsen map directly between 
LCSs and syntax, so there is no systematic link to background knowledge yet (which, 
as we have pointed out, would be useful for generation). Besides, as we mentioned in 
Section 3, LCS representations use primitives (BECOME, INCH, d), where we opt for 
a more fine-grained decomposition of the underlying event. 
For the alternations investigated, we have chosen the approach of defining a single 
base form from which alternated forms are derived. For other alternations, this might 
not be feasible or practical--in such cases, different lexical entries are to be used. There 
is, on the other hand, a line of research that questions the utility of distinguishing a 
base form from a more complex one in an alternation. For example, Saint-Dizier (1996) 
states that his approach to alternations deliberately avoids three difficulties: the need 
to define a basic form from which alternations are produced; the need to explain the 
relation between the basic form and the alternated one; and the need to account for 
changes in meaning produced by the alternation. It seems that the work presented 
in this paper aims precisely at those questions that Saint-Dizier's approach proposes 
to better leave aside. For generation, however, we believe that a system must know 
about the fine-grained changes in meaning that a verb alternation implies--a generator 
has to relate some semantic input representation to verb meaning, after all, and that 
includes alternations. And if the semantic change induced by an alternation can be 
described by a general rule that covers a whole class of verbs, a useful abstraction is 
gained. 
The final point to consider is the question of admitting lexical rules into one's 
framework. For example, Sanfilippo (1994) argues against this instrument on the 
grounds that there is no general control regime on lexical rules that would determin- 
istically restrict any polysemic expansion. Instead, he advocates coding the alternative 
lexical forms in a hierarchy of typed feature structures, where the underspecified forms 
subsume the specific ones. His criticism applies to the notion of rules that are trig- 
gered automatically and proceed to derive new forms without principled limitations. 
Our "defensive" approach of listing applicable rules in the lexical entries avoids this 
problem but at the same time raises the question of why rules should be preferable to 
a simple enumeration of forms. We return to this point in section 5.3. 
425 
Computational Linguistics Volume 24, Number 3 
5.2 Lexicalization in NLG 
In MOOSE, the lexicon is the central device for mapping between input representations 
and intermediate sentence-semantic representations. The idea of using the lexicon early 
in the generation process is not new; it has been realized in several other generators, for 
example in the frame-oriented system DIOGENES (Nirenburg and Nirenburg 1988). In 
contrast to earlier systems, however, MOOSE strengthens the role of lexical semantics 
in the generation process by distinguishing between the SitSpec and SemSpec levels 
and clearly specifying the relationships between the two (as done with the alterna- 
tion rules). Furthermore, we have emphasized that lexical choice should be seen as a 
constraint satisfaction process, similar to Reiter (1991), who focused his attention on 
nouns, while we have concentrated on verbs. 
There are several other generators using Penman as a front-end. For example, the 
DRAFTER system (Paris et al. 1995) builds SPLs and hands them over to Penman; 
contrary to MoosE, however, the domain model in DRAFTER is subsumed by the 
upper model, which significantly limits the range of lexical variation, as pointed out 
above. 
Working in the framework of systemic-functional grammar (SFG), both Wanner 
(1992) and Teich and Bateman (1994) employ SPL as an intermediate description, but 
they emphasize the integration of the SPL construction process into SFG. Wanner uses 
system networks to make fine-grained lexical choices in line with the three systemic 
metafunctions. Teich and Bateman develop system networks describing genre and 
register variation to drive the generation process, and they query an external domain 
model when building the SPL. In related work, Teich, Firzlaff, and Bateman (1994) 
present an implementation of Kunze's theory of semantic emphasis (cf. Section 4.3). 
From a "basic semantic scheme" annotated with emphasis labels, an SPL with ap- 
propriate roles and upper model concepts is constructed. The SPL can also contain 
an emphatic/nonemphatic feature, which might lead, for instance, to a dative shift. 
Hence, this work shares our interest in salience and indeed goes a step further than 
our present account in that the generation grammar can employ additional means for 
salience variation. However, in these three approaches, all lexical matters are taken 
to be part of the (huge) grammar processing the SPL. Thus, the central difference to 
the MOOSE approach is our step of promoting the lexicon to the crucial device for 
mapping between conceptual and sentence-semantic representations. We have argued 
that this step of keeping the lexicon separate and accessing it early has a number of 
advantages. 
Essentially the same difference holds between MOOSE and GOSSIP (Iordanskaja, 
Kittredge, and Polgu~re 1991), which also emphasizes the importance of lexical choice 
and paraphrasing abilities. Here, a powerful lexicalization mechanism is embedded 
in a meaning-text generation model following the theory of Mel'cuk, where lexical 
functions play a central role in mapping between different levels of representation. 
These are semantic and syntactic levels, though, whereas MOOSE focuses on the in- 
terrace between conceptual and semantic representations, and employs the lexicon at 
that point. 
Representations more similar to ours have been used by Dorr and Voss (1996), 
who employ Jackendoff's (1990) LCSs as an interlingua in machine translation, and by 
Di Eugenio (1993), who also represents LCS in a KL-ONE language but for purposes 
of analysis rather than generation. More specifically for NLG, structure mappings 
between fine-grained representations have been suggested for instance by Horacek 
(1990), Nogier and Zock (1992), and Nicolov, Mellish, and Ritchie (1996). In all these 
approaches, the input structure is directly mapped to a syntactic structure, though, 
while we have argued that an intermediate sentence-semantic level is advantageous 
426 
Stede Verb Alternations 
in order to explore generalizations (such as the alternation rules) as well as for multi- 
lingual purposes. 
5.3 Conclusions 
In a computational approach to the lexicon, word sense enumeration should not be 
the rule but be reserved for the exceptions (Pustejovsky 1995). In line with this view, 
our approach seeks to exploit generalizations by accounting for different forms of a 
verb with explicit alternation and extension rules that relate the changes in meaning to 
the changes in form. Ultimately, such an account establishes correspondences not only 
between different forms of the same verb but also between different verbs; for example, 
applying the causative extension to to rise yields (one form of) to raise. Interconnections 
of this kind have not yet been integrated into the system presented here, though. 
Three assumptions have guided the development of our account of verb alterna- 
tions: (1) There is a single base form from which other forms can be derived. (2) Alter- 
nation rules leave the denotation unchanged, and extension rules always add facets 
of meaning to the simpler denotation. (3) The changes in denotation correspond to 
changes in form, which can be characterized on a case-role level of description. In 
dealing with the telicity-related alternations discussed in Section 3, these assumptions 
have proven useful. For generalizing the approach to other alternations, assumption (2) 
could turn out to be too strong; in fact, even the causative extension might not always 
be monotonic when temporal adverbials are part of the sentence. In our framework, 
monotonicity is not a problem as long as the order of rule application is fixed anyway 
(cf. Figure 6). As soon as nonmonotonic rules are allowed, and the applicability of 
rules is no longer defined in the lexicon entries but triggered directly by the input, 
circularity is to be avoided: It needs to be ensured that rules reducing meaning reduce 
only parts that are not added by a different rule. 
Our selection of alternations was guided by their relationship to Aktionsart, in par- 
ticular to causation and telicity. Since the notion of Aktionsart is not a well-demarcated 
one in linguistics, and since the most comprehensive catalogue of alternations, the one 
by Levin (1993), has largely excluded Aktionsart-related problems, it is rather difficult 
to evaluate our approach in terms of "how many alternations" it covers. (Besides, 
we have argued in Section 3 that some of Levin's categorizations need refinement.) 
Clearly, there are other alternations involving telicity that we have not discussed here. 
Dorr and Olsen (1996) state that 27 of Levin's alternations add the telicity feature to a 
verb's meaning; many of these are rather specific and apply only to very few verbs. 
Among the more prominent ones are the unspecified object alternation (Tom ate~Tom 
ate a pizza) and the conative alternation (John cut at the bread~John cut the bread). Both 
lend themselves to extension rules as in our framework, because one form entails the 
other and adds information: When it holds that Tom ate a pizza, then it holds that 
Tom ate. Other alternations involve specific prepositions, such as Levin's through~with 
alternation: Alison pierced the needle through the cloth~Alison pierced the cloth with a needle. 
This does not pose problems for representing the changes in denotation, but renders a 
reliance on case roles--assumption (3) above---questionable; if suitable generalizations 
to similar prepositions cannot be found, the change in form ought to be stated directly 
on the syntactic level. 
Finally, we look at the question of evaluating our approach from the perspec- 
tive of natural language generation. From a descriptive viewpoint, as argued above, 
general lexical rules are to be preferred over enumerating word senses. Whether this 
preference also carries over to the design of practical NLG systems, however, merits 
some additional discussion. For the lexicalization step, we can either successively ap- 
ply alternation rules to a successfully matched base form, or compile out the various 
427 
Computational Linguistics Volume 24, Number 3 
alternated forms, which must then all be considered in matching against the input 
representation. While the first option obviously yields a much smaller lexicon, it is not 
self-evident whether it is faster or slower in a running system. 
As long as all alternated forms individually enter the matching phase, the compile- 
out option is hardly useful. Rather, compilation can be advantageous if only the most 
preferred form of the verb is considered first, and the other ones only upon request if 
the first did not work out. In this case, we are spared the effort of applying the rules 
to reach the desired form at run-time. Overall, the compilation decision hinges on the 
kind of criteria that the generator employs for its lexical choices. If the desired salience 
distribution is the central factor, then storing precompiled options and their salience 
information will be most effective. If considerations of lexical style lead to preferring 
one verb over a set of others irrespective of the specific alternation, then applying 
alternation rules only to the preferred verb will be more effective (in turn depending 
on how many similar verbs are ruled out and thus spared from the matching process). 
Thus, there appears to be no general answer; the size of the lexicon, including the 
ranges of nearly synonymous verbs, and the choice criteria used by the generator 
have to be taken into account. 
Acknowledgments 
The research reported in this paper 
originated at the University of Toronto 
(Canada) and at the Research Center for 
Applied Knowledge Processing (FAW) in 
Ulm (Germany). In the respective places, 
thanks to Graeme Hirst and Dietmar R6sner 
for discussions and advice, and to the 
Natural Sciences and Engineering Research 
Council of Canada (NSERC) and to FAW for 
financial support. I am grateful to the 
anonymous reviewers for their valuable 
suggestions for improving an earlier version 
of this paper. 
References 
Bach, Ernmon. 1986. The algebra of events. 
Linguistics and Philosophy 9. 
Bateman, John, Robert Kasper, Johanna 
Moore, and Richard Whitney. 1990. A 
general organization of knowledge for 
natural language processing: The penman 
upper model. Technical Report, USC/ISI, 
Marina del Rex CA. 
Brachrnan, Ronald and James G. Schmolze. 
1985. An overview of the KL-ONE 
knowledge representation system. 
Cognitive Science 9(2). 
Bussmann, Hadumod. 1983. Lexikon der 
Sprachwissenschaft. Kr6ner, Stuttgart. 
Di Eugenio, Barbara. 1993. Understanding 
Natural Language Instructions: A 
Computational Approach to Purpose Clauses. 
Ph.D. thesis, Department of Computer 
and Information Science, University of 
Pennsylvania. Technical Report 
MS-CIS-93-91. 
Dorr, Bonnie and Mari Broman Olsen. 1996. 
Multilingual generation: The role of 
telicity in lexical choice and syntactic 
realization. Machine Translation (11):37--47. 
Dorr, Bonnie and Claire Voss. 1996. A 
multi-level approach to interlingual MT: 
Defining the interface between 
representational languages. International 
Journal of Expert Systems 9(1). 
Fr6hlich, Marcel and R. van de Riet. 1997. 
Conceptual models as knowledge 
resources to text generation. In Proceedings 
of the Third International Workshop on the 
Application of Natural Language to 
Information Systems, Vancouver. 
Halliday, Michael. 1985. Introduction to 
Functional Grammar. Edward Arnold, 
London. 
Hasan, Ruqaiya. 1987. The grammarian's 
dream: Lexis as most delicate grammar. 
In Michael Halliday and Robin Fawcett, 
editors, New Developments in Systemic 
Linguistics. Volume 1. Pinter, London. 
Helbig, Gerhard and Wolfgang Schenkel. 
1973. W6rterbuch zur Valenz und 
Distribution deutscher Verben. VEB 
Verlag Enzyklop~idie, Leipzig. 
Horacek, Helmut. 1990. The architecture of 
a generation component in a complete 
natural language dialogue system. In 
Robert Dale, Chris Mellish, and Michael 
Zock, editors, Current Research in Natural 
Language Generation. Academic Press, 
London. 
Hov~ Eduard H. 1988. Generating Natural 
Language Under Pragmatic Constraints. 
Lawrence Erlbaum, Hillsdale, New Jersey. 
Iordanskaja, Lidia, Richard Kittredge, and 
Alain Polgu~re. 1991. Lexical selection 
and paraphrase in a meaning-text 
428 
Stede Verb Alternations 
generation model. In C4cile Paris, William 
Swartout, and William Mann, editors, 
Natural Language Generation in Artificial 
Intelligence and Computational Linguistics. 
Kluwer, Dordrecht. 
Jackendoff, Ray. 1990. Semantic Structures. 
MIT Press, Cambridge, MA. 
Kasper, Robert. 1989. A flexible interface for 
linking applications to Penman's sentence 
generator. In Proceedings of the DARPA 
Workshop on Speech and Natural Language 
Processing, University of Pennsylvania. 
Kunze, J6rgen. 1991. Kasusrelationen und 
semantische Emphase (studia grammatica 
XXXI). Akademie Verlag, Berfin. 
Levin, Beth. 1993. English Verb Classes and 
Alternations. University of Chicago Press, 
Chicago. 
Levin, Beth and Malka Rappaport Hovav. 
1995. Unaccusativity. MIT Press, 
Cambridge, MA. 
MacGregor, Robert and Robert Bates. 1987. 
The Loom Knowledge Representation 
Language. Techrfical Report 
ISI/RS-87-188, USC/ISI, Marina del Re)~ 
CA. 
Matthiessen, Christian and John Bateman. 
1991. Text Generation and Systemic 
Functional Linguistics: Experiences from 
English and Japanese. Pinter, London. 
Moens, Marc and Mark Steedman. 1988. 
Temporal ontology and temporal 
reference. Computational Linguistics 14(2). 
Nicolov, Nicolas, Chris Mellish, and Graerne 
Ritchie. 1996. Approximate generation 
from non-hierarchical representations. In 
Proceedings of the Eighth International 
Workshop on Natural Language Generation, 
Herstmonceux Castle, England. 
Nirenburg, Sergei and Irene Nirenburg. 
1988. A framework for lexical selection in 
natural language generation. In 
Proceedings of the 12 th International 
Conference on Computational Linguistics 
(COLING-88), Budapest. 
Nogier, Jean-Francois and Michael Zock. 
1992. Lexical choice by pattern matching. 
Knowledge Based Systems 5(3). 
Paris, C4cile, Keith Vander Linden, Markus 
Fischer, Anthony Hartle)~ Lyn Pemberton, 
Richard Power, and Donia Scott. 1995. A 
support tool for writing multilingual 
instructions. In Proceedings of the 
International Joint Conference on Artoqcial 
Intelligence (IJCAI-95), Montreal. 
Parsons, Terence. 1990. Events in the 
Semantics of English: A Study in Subatomic 
Semantics. MIT Press, Cambridge, MA. 
Pattabhiraman, T. 1992. Aspects of Salience in 
Natural Language Generation. Ph.D. thesis, 
School of Computing Science, Simon 
Fraser University. 
The Penman group. 1989. Unpublished 
Documentation of the Penman Sentence 
Generation System. USC/ISI, Marina del 
Rey, CA. 
Pinker, Steve. 1989. Learnability and 
Cognition: The Acquisition of Argument 
Structure. MIT Press, Cambridge, MA. 
Pustejovsky, James. 1991. The syntax of 
event structure. Cognition 41:47-81. 
Pustejovsk)~ James. 1995. The Generative 
Lexicon. MIT Press, Cambridge, MA. 
Reiter, Ehud. 1991. A new model of lexical 
choice for nouns. Computational Intelligence 
7:240-251. 
R6sner, Dietmar and Manfred Stede. 1994. 
Generating multilingual documents from 
a knowledge base: The TECHDOC 
project. In Proceedings of the International 
Conference on Computational Linguistics 
(COLING-94), Kyoto. 
Saint-Dizier, Patrick. 1996. Verb semantic 
classes based on 'alternations' and on 
WordNet-like semantic criteria: A 
powerful convergence. In Proceedings of the 
Workshop on Predicative Forms in Lexical 
Semantics and Lexical Knowledge Bases, 
Toulouse. 
Sanfilippo, Antonio. 1994. Word knowledge 
acquisition, lexicon construction, and 
dictionary compilation. In Proceedings of 
the International Conference on Computational 
Linguistics (COLING-94), Kyoto. 
Somers, Harold. 1987. Valency and Case in 
Computational Linguistics. Edinburgh 
University Press, Edinburgh. 
Stede, Manfred. 1996a. Lexical Semantics and 
Knowledge Representation in Multilingual 
Sentence Generation. Ph.D. thesis, Dept. of 
Computer Science, University of Toronto. 
Technical Report CSRI-347. 
Stede, Manfred. 1996b. Lexical paraphrases 
in multilingual sentence generation. 
Machine Translation 11:75-107. 
Stede, Manfred and Brigitte Grote. 1995. The 
lexicon: Bridge between language-neutral 
and language-specific representations. In 
Working notes of the IJCAI Workshop on 
Multilingual Text Generation, Montreal. 
Talmy, Leonard. 1988. The relation of 
grammar to cognition. In 
B. Rudzka-Ostyn, editor, Topics in Cognitive 
Linguistics. John Benjamins, Amsterdam. 
Teich, Elke and John Bateman. 1994. 
Towards the application of text 
generation in an integrated publication 
system. In Proceedings of the Seventh 
International Workshop on Natural Language 
Generation, Kennebunkport, ME. 
Teich, Elke, Beate Firzlaff, and John 
Bateman. 1994. Emphatic generation: 
429 
Computational Linguistics Volume 24, Number 3 
Employing the theory of semantic 
emphasis for text generation. Paper 
presented at the International Conference 
on Computational Linguistics 
(COLING-94), Kyoto. 
Tesni~re, L. 1959. t~le'ments de syntaxe 
structurale. Klincksieck, Paris. 
Vendler, Zeno. 1967. Linguistics and 
Philosophy. Cornell University Press, 
Ithaca, New York. 
Wanner, Leo. 1992. Lexical choice and the 
organization of lexical resources in text 
generation. In Proceedings of the European 
Conference on Artificial Intelligence (ECAI92), 
Vienna. 
White, Michael. 1994. A Computational 
Approach to Aspectual Composition. Ph.D. 
thesis, Dept. of Computer and 
Information Science, University of 
Pennsylvania. 
430 
