Extracting Semantic Features for Aspectual Meanings ... - CiteSeerX

0 downloads 0 Views 184KB Size Report
May 5, 1994 - modules dealing with time, aspect, space etc. This means that we do .... x-times yesterday sing-def telic-action simple non-perfect past sing- def.
Extracting Semantic Features for Aspectual Meanings from a Syntactic Representation Using Neural Networks Gabriele Scheler Institut fur Informatik Technische Universitat Munchen, D-80290 Munchen e-mail: [email protected] May 5, 1994 Abstract

The main point of this paper is to show how we can extract semantic features, describing aspectual meanings, from a syntactic representation. The goal is to translate English to Russian aspectual categories. This is realized by a specialized language processing module, which is based on the concept of vertical modularity. The results of supervised learning of syntactic-semantic correspondences using standard back-propagation show that both learning and generalization to new patterns is successful. Furthermore, the correct generation of Russian aspect from the automatically created semantic representations is demonstrated.

1 Introduction A common goal of theoretical linguistic work as well as machine translation research is the construction of semantic representations, which can be used as interlingual representations in the context of machine translation. Constructing semantic representations is often controversial and results are dicult to evaluate, because little is known on what constitutes a semantic representation in the human brain and which information it must contain. However, semantic representations are a useful concept, when we have a speci c language processing task. In di erent tasks, we may use the same or similar semantic representations to interface with di erent modules for further processing. In this work, we concentrate on the task of translation, and accordingly interface the semantic representation with a language generation 1

module for another language. This task-oriented approach has the important advantage that the speci c contribution of a semantic representation can be assessed and valued. The main point of this paper is to show how we can extract semantic features, describing aspectual meanings, from a specialized syntactic representation. The overall goal is to translate English to Russian aspectual categories, which can not be handled well by rule-based machine translation. The reason is mainly that the constellations of contextual information that determine semantic content and accordingly morphological categories in another language are too numerous and too diverse to be used as triggers for rules. In this paper, emphasis is put on the interpretation process. Generation of the correct aspect in Russian from the semantic representation is achieved by a learned pattern classi er, which has been described in [Scheler, 1994a]. source text

?

specialized syntactic representation encoding

?

binary source text representation set of aspectual meanings

encoding

-

?

feature extraction

binary feature vector

?

pattern classi cation

language-speci c morphological category assignment

?

target text

Figure 1: Model of the Translation Process In Figure 1 a model of the complete translation process is shown. Semantic feature extraction is realized by a specialized language processing module for dealing with aspectual information. This vertical modularity has been inspired by neurolinguistic research ([McCarthy and Warrington, 1988],[Marshall, 1988]). Instead of the more usual horizontal modules for morphology, syntax, lexical and sentential semantics, we use vertical modules that can be de ned as areas of processing that combine e.g. temporal morphology, temporal adverbials, temporal 2

semantics and temporal cognition. Thus we have a number of specialized modules dealing with time, aspect, space etc. This means that we do not use a full syntactic representation, with all information that can be extracted by a parsing mechanism, but a specialized syntactic representation, containing in this case quanti cation of noun phrases (subject and direct object), lexical category of the verb with respect to aspectual properties, and temporal and modal sentential adverbs. There is quite a number of work which is related to the present one. Feature-based syntactic representations for pattern association with semantic representations have been used by [McClelland and Kawamoto, 1986] and [John and McClelland, 1990]. Their models are used to deal with the association between syntactic cases and semantic roles. The semantic representations achieved were not interfaced with any other module. A combined interpretation and generation approach has been used by [Miikkulainen, 1990] for the analysis of relative clauses. There, the internal representation also consisted of situations and roles. Another vertical module was modeled by [Munro and Tabasko, 1991]. Here locative prepositions were interpreted as spatial relations. However the model lacked generalization in the creation of semantic representations. The paper is structured as follows: In the next section, the notion of specialized syntactic representation is further explained, and the grammar that has been used in setting up the representation is presented. The following sections contain results of experiments in syntactic-semantic association and in generation. In the conclusion, results are assessed and further questions are being raised.

2 Syntactic features for aspectual meanings The idea in using a specialized syntactic representation is that linguistic input is fed into several specialized modules for analysis, which take the necessary information for their task and retransform it into a suitable form, probably in several steps. As a starting point in this work we use a hypothetical intermediate representation, which can be gained from free text by a fairly simple tagger or parser. Relevant syntactic information has been selected in accordance with the detailed presentation in [Scheler, 1984], cf. [Partee, 1973; Vendler, 1967]. The categorizations presented here are all xed lexical decisions and not contingent on the use of these words in a given sentence. This is in accordance with the principle of a unidirectional parser, and could be changed by the integration of this semantic feature extraction module into a connectionist parser. The syntactic representation consists of eight slots, i.e. placeholders for functionally di erent terms, and a number of values for each 3

slot. The following description is the \grammar" of the specialized syntactic representation for aspectual meanings:

quanti cation of subject: singular definite,

plural definite, singular indefinite and plural indefinite, as they are indicated by mor-

phological marking on English noun phrases. temporal adverbials: These are distinguished into classes according to the preposition (at, for, on) (or conjunction (when)) used, and the type of noun that forms the head, such as point (\8 o'clock") or period (\three hours"). Words such as now, today, yesterday form a category of their own.

morphological verb markers: past/present, perfect/non-perfect, progressive/simple.

sentential adverbs: These comprise   



adverbs always, sometimes, x-times (\ ve times"), usually adverbs such as suddenly, just, as independent categories, adverbials of manner (\slowly", \with great care", \well"), which have been categorized into those indicating processual nature mannerproc, and those that do not and are neutral in this regard (manner), and sentential negation (negation).

quanti cation of direct object: singular definite, plural definite, singular indefinite and plural indefinite.

lexical verb classes: A list of verb-classes sucient for the chosen examples has been derived from [Scheler, 1984]: telic-action, achieve-

ment, active-perception, atelic-action, passive-perception, gradual state-change, stages-of-actions (\try, begin, nish"), state, motion

This syntactic representation is transformed into a semantic feature representation with pattern association techniques. The semantic representation consists of fteen features like habituality, event-extension, eventtype, degree of completion, duration and between two and ve values for each of these features, such as for event-type: state, atelic event, telic event, cause-state (cf. appendix A for a complete list, cf. also [Scheler, 1994a]). These are the basis for the selection of an aspectual category in another language. A total of 49 patterns was created. The examples are given in appendix B. Initially, about 40 patterns were selected from a standard English grammar 4

([Thompson and Martinet, 1969]) exemplifying the di erent uses of English tenses and aspects. However, it turned out that the results of this approach could be much enhanced, when several syntactic variants for the most salient examples were created. I.e. examples were created to allow the classi er to determine the in uence of minor syntactic changes on the semantic representations. Five variations on the sentence \He is always doing homework" are shown in Table 2, together with their syntactic and proposed semantic representation. This is a method which is well known in theoretical linguistic He has not done his homework.

negation * sing-def telic-action simple perfect present sing-def

past * * relational-to-present holistic * * * action existential telic negated * single non-habitual. He was doing homework from 5 to 7.

* from-point-to-point sing-def telic-action prog non-perfect past plu-indef

past * * non-relational processual * occurs-at-period-in * action referential atelic * * single non-habitual. Sometimes he is doing his homework with much care.

manner-proc,sometimes * sing-def telic-action prog non-perfect present sing-def

* * non-relational processual * * * action existential atelic * long-duration single habitual. Right now all students are doing their homework. * now plu-def telic-action prog non-perfect present sing-def

present * * non-relational processual extends-around * * action referential atelic halfway-through * single non-habitual. Yesterday, Tom did his homework twice.

x-times yesterday sing-def telic-action simple non-perfect past singdef

past hesternal * non-relational holistic * occurs-at-period-in * action referential telic completed * repeated non-habitual.

Table 1: Training Patterns with Syntactic Variation work (cf. e.g. [Verkuyl, 1993] on aspectology), and which is also used much in second-language learning, i.e. the attempt to implement a speci c linguistic system onto an existing one. Possibly it is being used in rst language acquisition as well.

5

3 Results of experiments The syntactic representations, consisting of six slots with between two and thirteen di erent values, have been translated into binary feature vectors, using the following encoding technique: For each slot or feature, it was determined how many bits are necessary to code all possible values for this feature. A neutral value ('') has been added for each feature. The resulting coding has a length of 25 bits. The same technique has been applied to the semantic representation, which has 15 features, resulting in 34 bits. The architecture was that of a feedforward neural network, trained with standard backpropagation (cf. [Zell and others, 1993]). A separate encoding level for the syntactic representation was included, i.e. a chance to recode the slot-value representation in an opaque way, which proved an asset in generalization (cf. [Hinton, 1986] for a discussion of recoding). The number of units in the di erent layers were 25-15-12-34. The training technique used was cross-validation by leaving-one-out (cf. [Weiss and Kulikowski, 1991], pp. 26{39) which is a preferential method for small samples of less than 100 patterns. Because semantic representations for the training set have to be created by hand, it is dicult to obtain large samples. However, as we shall see, the network learns well with small samples and can then generate additional semantic representations, which can be checked for their suitability by interfacing it with another task. The network was trained with 48 patterns (2500 cycles) and then the remaining pattern was tested for generalization. A typical result for learning is shown in Table 2. This process was repeated 49 times, until all patterns had been tested. The results of generalization are also shown in Table 2. The third row (marked by y) indicates the results, when we exclude \outliers", i.e. patterns with more than 10 errors. No. patterns No. patterns with n errors Avg. No. Phase total correct (%) n  5 (%) n > 5 (%) errors/pattern learning 48 42 (87.5%) 48 (100%) 0 (0%) 0.18 generalization 49 5 (10.2 %) 35 (71.4%) 14 (28.6%) 3.81 generalizationy 46 5 (10.8 %) 35 (76.1%) 11 (23.9%) 3.23 Table 2: Learning and Generalizing Syntactic-to-Semantic Pattern Association The results show that learning, i.e. implementing the functional relationship is no problem, as expected with a small training set. The gures on generalization show that some implicit rules on how to set semantic values given syntactic input have been abstracted. The task was not an easy one, and an average of three to four errors per pattern remained. 6

The results can be discussed as follows: Most information that is needed in order to determine semantic features for aspectual meanings is indeed local syntactic and has been captured by the specialized syntactic representation. However, for most patterns certain features could not be uniquely determined. There are several strategies possible for a remedy: One possibility would be to include more syntactic information. However, a careful analysis has shown that further information is usually drawn from the same set of aspectually relevant syntactic features ([Scheler, 1984]). To give an example: the present perfect of recency, (proximity: recent) is usually indicated by the adverb just. When 'just' is missing, it is possible that a pattern nonetheless is somehow \similar" to other typical present perfects of recency. However, one would expect this similarity to be expressed by the same lexical category of the verb (achievement or telic-action) and/or by the absence of a direct object, typical present perfects of recency being of the type \He has (just) gone", \I have (just) eaten" etc. Therefore one would not expect to gain much by including more local syntactic information. Another question is whether features of lexical content and situational context would improve performance. It is dicult to obtain reliable features describing situational context (but cf. [Gallant, 1991] for an attempt). Lexical content features would have to be imported from other analyzing modules, a problem that has not been satisfactorily solved (cf. [McClelland et al., 1989], [Scheler, 1989] for a proposal on lexical-syntactic interaction during processing). An analysis of the example sentences that were not well generalized would lead to expect some improvement from the inclusion of a wider semantic context. The third idea is to have the system modify the semantic representation from knowledge of the likelihood of semantic representations. Such knowledge could be incorporated into weights in lateral connections among the semantic feature-values. To make a distinction between such knowledge from experience and the learning related to the present task requires a di erent learning method. In this way, performance would be probably signi cantly improved.

4 Generation of aspectual categories The second experiment concerns the generation of aspectual categories from the semantic representations achieved by generalization, i.e. semantic representations that the system has found on the basis of a training set. The training set was rst manually translated from English to Russian1 to provide a measure of success for the automatic translation. Then a network was trained to classify semantic representations according to Russian 1I

wish to thank Dr. Terterjan for carefully checking all translations.

7

aspect. Finally, the generalized semantic representations from the previous experiment (Section 3) were given to the Russian aspect classi er. The results are given in Table 3. They show that the errors were negligible in determining Russian aspect. The explanation for this successful performance is straightforward: The average error for generalization was less than 4 bits out of 34 bits. In addition, when neighboring binary features are wrong, this is often only one error on the symbolic level. Accordingly of the 15 features to determine a binary classi cation, only 1 or 2 were set di erently than in the learning task. Two of the misgenerated patterns were also \outliers" in the interpretation task. No. patterns Phase total correct(%) learning 48 47 (97.9%) generalization 49 45 (91.8%) Table 3: Generating Russian Aspect from System-Created Semantic Representations

5 Conclusion It has been shown that extracting semantic features for a speci c area of meaning and generation of morphological categories is a task that is well suited for neural network models. Standard back-propagation is a useful method to determine whether the problem speci cation is correct, i.e. whether there are generalizable patterns involved and whether the problem has been put in such a way, as to make a functional implementation feasible. For a closer look, we want to use more biologically inspired learning mechanisms, and also simpler mechanisms (cf. [Scheler, 1994b]). In this paper, however, we are concerned with problem speci cation and a preliminary functional implementation. We are successful here, if the proposed function is learnable and generalizable by a standard mechanism. On the other hand, this approach is a vast improvement over rule-based methods, because the in uence of diverse syntactic information onto the semantic representation has not to be manually analyzed in detail in order to set up a great number of individual rules with di erent trigger patterns. Instead, the various aspects of the information in the syntactic representation are extracted and recast into a form which corresponds to the tasks of whatever we want to do with the information. Generation is modelled as a pattern classi cation process, where a morphological selection is made on the basis of a semantic feature representation. Here the advantage of a 8

fault-tolerant mechanism is evidenced by the fact that generation results are highly successful even from semantic representations containing errors. An important goal for further research is the translation of aspectual forms from continuous texts. For this goal, a syntactic tagger has to be added to the system. A similar system can then be used as a grammar checker for the correctness of aspectual forms.

A Semantic Representation: Features and Values

1. event-time past - present - future (3) 2. proximity (past) recent - hodiernal - hesternal - less than a year - more than a year (5) 3. proximity (for futuric action) immediately - not immediate (2) 4. relational relational to past - relational to present - non relational (3) 5. event-extension punctual (instantaneous) - extended (processual,durative) - holistic (3) 6. reference point in time (event:) occurs at - starts at - extends around (3) 7. reference period (event:) occurs at point in - occurs at period in - occurs at starting point - occurs at end point (4) 8. reference times/periods (event occurs at:) all of them (all/most) - some of them (2) 9. action-status action - non action(2) 10. reference-type existential (inde nite) - referential (de nite) - general (mass noun reference) (3) 11. event-type state - atelic event - telic event - cause state (4) 9

12. degree of completion attempt - halfway through - completed - negated (4) 13. duration long duration - limited duration (short duration) (2) 14. number of occurrences single - repeated (2) 15. habituality habitual - non habitual (2)

B Example sentences 1. At six o'clock I am usually bathing the baby. usually at point sing def telic-action prog non perfect present sing def * * non relational processual extends around * * action existential telic halfway through * single habitual. 2. Tom is always going away for the weekends. always for period sing def achievement prog non perfect present * present * * non relational punctual * occurs at starting point * action existential cause state completed * single habitual. 3. He is always doing homework. always * sing def telic-action prog non perfect present mass present * * non relational processual * * * action existential telic halfway through long duration single habitual. 4. I am always tripping over this suitcase. always * sing def achievement prog non perfect present sing def present * * non relational punctual * * * action referential cause state halfway through * repeated non habitual. 5. I taste salt in my porridge. * sing def state simple non perfect present mass present * * non relational holistic * * * non action referential state * * * non habitual. 6. I hear you well. manner * sing def state simple non perfect present sing def present * * non relational holistic * * * non action referential state * * * non habitual. 7. Their children are really very quiet. * plu def state simple non perfect present attr * * non relational processual * * * non action general state * * * non habitual. 8. I can't open the door, I am having a bath. now sing def atelic-action prog non perfect present sing indef 10

present * * non relational processual extends around * * action referential telic halfway through * single non habitual. 9. Are you liking this excursion? No I'm hating it. now sing def state prog non perfect present sing def present * * non relational processual extends around * * non action referential state halfway through long duration single non habitual. 10. I don't expect much of him. negation * sing def state simple non perfect present mass present * * non relational processual * * * non action general state * * * non habitual. 11. When John entered, I was bathing the baby. when point sing def telic-action prog non perfect past sing def past * * non relational processual extends around * * action existential telic halfway through * single non habitual. 12. I bathe the baby every day. every unit of time sing def telic-action simple non perfect present sing def * * non relational holistic * * all periods action existential telic completed * single habitual. 13. On Sundays I bathe the baby. on periods sing def telic-action simple non perfect present sing def * * non relational holistic * * all periods action existential telic completed * single habitual. 14. Most people bathe their baby every day. every unit of time plu def telic-action simple non perfect present sing def * * non relational holistic * * all periods action general telic completed * single habitual. 15. Tom is always going away for the weekends. always for period sing def achievement prog non perfect present * present * * non relational punctual * occurs at starting point * action existential cause state completed * single habitual. 16. Today Tom is going away to visit his brother. today sing def achievement prog non perfect present to do future * immediately non relational punctual * occurs at point in * action referential cause state completed * single non habitual. 17. When Mary entered, Tom went away. when point sing def achievement simple non perfect past * past * * non relational punctual occurs at * * action referential cause state completed * single non habitual. 18. Tom has just gone away to visit his brother. just * sing def achievement simple perfect present to do past recent * relational to present punctual * * * action existential cause state completed * single non habitual. 11

19. Tom has just gone away . just * sing def achievement simple perfect present * past recent * relational to present punctual * * * action existential cause state completed * single non habitual. 20. He always does his homework. always * sing def telic-action simple non perfect present sing def present * * non relational processual * * * action existential telic completed * single habitual. 21. He has not done his homework. negation * sing def telic-action simple perfect present sing def past * * relational to present holistic * * * action existential telic negated * single non habitual. 22. He was doing homework from 5 to 7. from point to point sing def telic-action prog non perfect past mass past * * non relational processual * occurs at period in * action referential telic * * single non habitual. 23. Sometimes he is doing his homework with much care. manner long,sometimes * sing def telic-action prog non perfect present sing def * * non relational processual * * * action existential telic * long duration single habitual. 24. Right now all students are doing their homework. now plu def telic-action prog non perfect present sing def present * * non relational processual extends around * * action referential telic halfway through * single non habitual. 25. Yesterday, Tom did his homework twice. x times yesterday sing def telic-action simple non perfect past sing def past yesterday * non relational holistic * occurs at period in * action referential telic completed * repeated non habitual. 26. In Montreal at the airport I tripped over my suitcase and sprained an ankle. in place sing def achievement simple non perfect past sing def past longer * non relational punctual * occurs at point in * action referential cause state completed * single non habitual. 27. Suddenly they tripped over the suitcase and fell. suddenly * plu def achievement simple non perfect past sing def past * * non relational punctual occurs at * * action referential cause state completed * repeated non habitual. 28. I am glad I haven't tripped over it. negation * sing def achievement simple perfect present sing def past recent * relational to present punctual * * * action existential cause state negated * single non habitual. 12

29. Unfortunately I am tripping over suitcases everywhere. every unit of time sing def achievement prog non perfect present mass * * non relational punctual * occurs at point in all periods action general cause state completed * single habitual. 30. For months I was tasting salt in my porridge, before I got to know, why. for period sing def state prog non perfect past mass past longer * non relational punctual * occurs at point in * non action existential state * * * habitual. 31. Today I tasted salt in my porridge. today sing def state simple non perfect past mass past hodiernal * non relational punctual * occurs at point in * non action referential state completed * single non habitual. 32. I was tasting the porridge, when it exploded with a bang. when point sing def active-perception prog non perfect past sing def past * * non relational processual extends around * * action referential telic halfway through * single non habitual. 33. I have tasted the porridge thoroughly, it tastes good. manner long * sing def active-perception simple perfect present sing def past * * relational to present holistic * * * action existential telic completed long duration single non habitual. 34. I've been hearing all about this accident from him. * sing def state prog perfect present all of x past recent * relational to present processual * * * non action existential state completed long duration single non habitual. 35. Suddenly I heard an explosion. suddenly * sing def state simple non perfect past sing indef past * * non relational punctual * * * non action referential atelic completed * single non habitual. 36. For a long time we were hearing little explosions. for period plu def state prog non perfect past mass past * * non relational processual * occurs at period in * non action referential atelic * long duration repeated non habitual. 37. In winter birds hear a lot better than in summer. manner in period plu indef state simple non perfect present * * * non relational holistic * occurs at period in some periods non action general state * * * habitual. 38. The children are being very quiet. now plu def active-perception prog non perfect present attr present * * non relational processual * * * non action referential state halfway through long duration single non habitual. 39. On sundays, the children are usually very quiet. 13

usually on periods plu def active-perception simple non perfect present attr * * non relational holistic * occurs at period in * non action existential state * * * habitual. 40. On sundays, the children were usually very quiet. usually on periods plu def active-perception simple non perfect past attr past * * non relational holistic * occurs at period in * non action existential state * * * habitual. 41. On sundays, children are usually very quiet. usually on periods plu indef active-perception simple non perfect present attr * * non relational holistic * occurs at period in * non action general state * * * habitual. 42. At 8 he was having breakfast. at point sing def atelic-action prog non perfect past * past * * non relational processual extends around * * action referential telic halfway through * single non habitual. 43. He was always having breakfast at 8 in the morning. always at point sing def atelic-action prog non perfect past * past * * non relational processual extends around * * action existential telic halfway through * single habitual. 44. Every morning, from 8 to 8.30 I am having a bath. every unit of time,from point to point sing def atelic-action prog non perfect present * * * non relational processual * occurs at period in * action existential telic halfway through * single habitual. 45. When the door rang, we were having breakfast in the kitchen. when point plu def atelic-action prog non perfect past in place past * * non relational processual extends around * * action referential telic halfway through * single non habitual. 46. I have just had a bath. just * sing def atelic-action simple perfect present * past recent * relational to present holistic * * * action existential telic completed * single non habitual. 47.I have had a bath twice today. x times today sing def atelic-action simple perfect present * past recent * relational to present holistic * * * action existential telic completed * repeated non habitual. 48. I hate excursions. * sing def active-perception simple non perfect present mass * * non relational processual * * * non action general state * * * non habitual. 49. I am expecting a letter today. today sing def active-perception prog non perfect present sing indef present * * non relational processual * occurs at period in * non action referential state halfway through * single non habitual. 14

References

[Gallant, 1991] S.I. Gallant. A practical approach for representing context and for performing word sense disambiguation using neural networks. Neural Computation 3(3), pages 293{309, 1991. [Hinton, 1986] G.E. Hinton. Learning distributed representations of concepts. In Proceedings of the Eighth Annual Conference of the Cognitive Science Society, 1986. [John and McClelland, 1990] M. F. St. John and J. L. McClelland. Learning and applying contextual constraints in sentence comprehension. Arti cial Intelligence, 46(1-2):217{257, 1990. [Marshall, 1988] John C. Marshall. Sensation and semantics. Nature, 334:378, August 1988. [McCarthy and Warrington, 1988] R. McCarthy and E. Warrington. Evidence for modality-speci c meaning systems in the brain. Nature 334, pages 428{430, 1988. [McClelland and Kawamoto, 1986] J. McClelland and A. Kawamoto. Mechanisms of sentence processing: Assigning roles to constituents. In D. E. Rumelhart and J. L. McClelland, editors, Parallel distributed processing: Explorations in the microstructure of cognition, pages 77{109. Cambridge, MA: MIT Press, 1986. [McClelland et al., 1989] J.L. McClelland, Mark St.John, and Roman Taraban. Sentence comprehension: A parallel distributed processing approach. Language and Cognitive Processes, 4(3/4):287{335, 1989. [Miikkulainen, 1990] Risto Miikkulainen. A PDP architecture for processing sentences with relative clauses. In Proceedings of Coling'90, 1990. [Munro and Tabasko, 1991] P. Munro and M. Tabasko. Translating locative prepositions. In Proceedings of NIPS-91, volume 3, pages 598{604, 1991. [Partee, 1973] Barbara Partee. Some Analogies between Pronouns and Temporal Expressions in English. Journal of Philosophy, 70:601{609, 1973. [Scheler, 1984] Gabriele Scheler. Zur Semantik von Tempus und Aspekt, insbesondere des Russischen. Master's thesis, LMU, Munchen, April 1984. 15

[Scheler, 1989] Gabriele Scheler. Remarks concerning the interaction of grammar and semantics. Technical report, Computerlinguistik, Universitat Heidelberg, 1989. [Scheler, 1994a] Gabriele Scheler. Multilingual generation of grammatical categories. Technical Report FKI-190-94, Technische Universitat Munchen, April 1994. [Scheler, 1994b] Gabriele Scheler. Pattern classi cation with adaptive distance measures. Technical Report FKI-188-94, Technische Universitat Munchen, January 1994. [Thompson and Martinet, 1969] A.J. Thompson and A.V. Martinet. A Practical English Grammar. Oxford University Press, 1969. [Vendler, 1967] Zeno Vendler. Linguistics in Philosophy. Cornell University Press, 1967. [Verkuyl, 1993] Henk J. Verkuyl. A theory of aspectuality. The interaction between temporal and atemporal structure. Cambridge Studies in Linguistics 64. Cambridge University Press, 1993. [Weiss and Kulikowski, 1991] Sholom M. Weiss and Casimir A. Kulikowski. Computer Systems That Learn. Morgan Kaufmann, 1991. [Zell and others, 1993] Andreas Zell et al. Snns User Manual v. 3.1. Universitat Stuttgart: Institute for parallel and distributed high-performance systems, 1993.

16