<?xml version="1.0" standalone="yes"?>
<Paper uid="J98-3003">
  <Title>A Generative Perspective on Verb Alternations</Title>
  <Section position="7" start_page="423" end_page="427" type="concl">
    <SectionTitle>
5. Summary and Related Work
</SectionTitle>
    <Paragraph position="0"> We have proposed an NLG framework, together with suitable representation schemes, that can systematically produce a range of verb alternations from a common underlying 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).</Paragraph>
    <Paragraph position="1"> For generation, our approach uses two distinct ontologies: a language-neutral domain 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 conceptual 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 implemented 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.</Paragraph>
    <Paragraph position="2"> The examples discussed in this paper (including the alternated forms) were generated 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 depends on quite fine-grained representations in both domain model and lexicon, it is difficult to make statements on how well the approach &amp;quot;scales up.&amp;quot; Large-scale (automatic) 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.</Paragraph>
    <Section position="1" start_page="424" end_page="424" type="sub_section">
      <SectionTitle>
Stede Verb Alternations
</SectionTitle>
      <Paragraph position="0"> In the following, we compare our approach to some related work on verb alternations and on lexicalization in NLG. Finally, we draw some conclusions as to the overall scope of the work and its utility for NLG.</Paragraph>
    </Section>
    <Section position="2" start_page="424" end_page="425" type="sub_section">
      <SectionTitle>
5.1 Alternations
</SectionTitle>
      <Paragraph position="0"> 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.</Paragraph>
      <Paragraph position="1"> 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 representations 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 proposed 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 denotation 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.</Paragraph>
      <Paragraph position="2"> 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.</Paragraph>
      <Paragraph position="3"> 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 deterministically 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 triggered automatically and proceed to derive new forms without principled limitations.</Paragraph>
      <Paragraph position="4"> Our &amp;quot;defensive&amp;quot; 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.</Paragraph>
      <Paragraph position="5">  Computational Linguistics Volume 24, Number 3</Paragraph>
    </Section>
    <Section position="3" start_page="425" end_page="426" type="sub_section">
      <SectionTitle>
5.2 Lexicalization in NLG
</SectionTitle>
      <Paragraph position="0"> 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 alternation 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.</Paragraph>
      <Paragraph position="1"> 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.</Paragraph>
      <Paragraph position="2"> 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).</Paragraph>
      <Paragraph position="3"> From a &amp;quot;basic semantic scheme&amp;quot; annotated with emphasis labels, an SPL with appropriate 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.</Paragraph>
      <Paragraph position="4"> 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.</Paragraph>
      <Paragraph position="5"> 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.</Paragraph>
      <Paragraph position="6"> These are semantic and syntactic levels, though, whereas MOOSE focuses on the interrace between conceptual and semantic representations, and employs the lexicon at that point.</Paragraph>
      <Paragraph position="7"> 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</Paragraph>
    </Section>
    <Section position="4" start_page="426" end_page="426" type="sub_section">
      <SectionTitle>
Stede Verb Alternations
</SectionTitle>
      <Paragraph position="0"> in order to explore generalizations (such as the alternation rules) as well as for multi-lingual purposes.</Paragraph>
    </Section>
    <Section position="5" start_page="426" end_page="427" type="sub_section">
      <SectionTitle>
5.3 Conclusions
</SectionTitle>
      <Paragraph position="0"> 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.</Paragraph>
      <Paragraph position="1"> Three assumptions have guided the development of our account of verb alternations: (1) There is a single base form from which other forms can be derived. (2) Alternation 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.</Paragraph>
      <Paragraph position="2"> Our selection of alternations was guided by their relationship to Aktionsart, in particular 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 &amp;quot;how many alternations&amp;quot; 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.</Paragraph>
      <Paragraph position="3"> 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.</Paragraph>
      <Paragraph position="4"> 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.</Paragraph>
      <Paragraph position="5"> Finally, we look at the question of evaluating our approach from the perspective 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 apply alternation rules to a successfully matched base form, or compile out the various  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.</Paragraph>
      <Paragraph position="6"> As long as all alternated forms individually enter the matching phase, the compileout 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).</Paragraph>
      <Paragraph position="7"> 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.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>