Towards a Uniform Treatment of Swedish Verb ... - Semantic Scholar

9 downloads 0 Views 169KB Size Report
uniformly by a “rule schema” approach to grammar in a unification based framework. The virtues of the ... Såg han Maria? “Did he see Mary?” Vem såg han?
Towards a Uniform Treatment of Swedish Verb Syntax and Semantics Bjšrn GambŠck

Natural Language Processing Group Swedish Institute of Computer Science Box 1263, S - 164 28 Kista, Stockholm, Sweden e-mail: [email protected]; telephone: +46 Ð 8 Ð 752 15 05 Abstract The paper discusses how different types of Swedish verbs can be treated uniformly by a Òrule schemaÓ approach to grammar in a unification based framework. The virtues of the method are exemplified at the syntactic and lexical levels, while the paper shows that some problems still occur in the semantics, mainly when aiming for a treatment which includes all types of auxiliaries. The theory outlined here incorporates the semantic information added by for example tense and modal auxiliaries as specific parameters in the verb semantics and relies on complicated subcategorization schemes to be treated by the lexicon, thus keeping the grammar rules as simple as possible. 1. Introduction From a theoretical view-point, the aim of the work described in the present paper is to establish a uniform treatment of Swedish verb-phrases of any kind, be it with or without modification or with different types of verbal complements. Reaching that goal, a unification based grammar theory has been essential, but before going into details of the theory, this first section will start out by describing the framework in which the theory has been built. The rest of the paper is then laid out as follows: Section 2 discusses the syntactic treatment of verb-phrase formation as well as different types of Swedish verbs. The semantic part of the treatment is addressed in Section 3, while Section 4 deals with negation Ð a subject which will be excluded from the discussion in the first sections. Section 5, finally, sums up the paper and points to areas where further improvements may be possible. The grammar treatment outlined in the paper has to quite a large extent been influenced by the fact that it was going to be implemented in a actual computer system, in the Swedish grammar used in an English-to-Swedish speech-tospeech translation prototype called the Spoken Language Translator, SLT

(AgnŠs et al 1994). A prototype developed jointly by SICS, SRI International and Telia Research. The system on the Swedish side, the Swedish Core Language Engine (S-CLE, GambŠck & Rayner, 1992) is a general-purpose natural-language processing system developed from its English counter-part, the SRI Core Language Engine (CLE, Alshawi, 1992). The grammar formalism used in both systems is a feature-category unification type with declarative bidirectional rules, that is, the grammar can be used both for language analysis and for generation (it is just compiled in different ways depending on the direction it is to be used). A natural-language sentence that is input to the S-CLE is analysed to a logicalform like representation called QLF, Quasi-Logical Form, a conservative representation of the meaning of an input sentence based on purely linguistic evidence. Deriving a QLF from an NL-sentence involves the processing steps shown in Figure 1: first morphological analysis locate the correct word-senses and inflected forms of the input string, then syntactic parsing and (compositional) semantic analysis derive the parse tree(s) and its corresponding QLF-representation. Later processing steps (e.g., reference resolution and quantifier scoping) will try to further instantiate the QLF, aiming at deriving a ÒtrueÓ logical form (that is, a context and application dependent representation). NL

Morphology

Syntax

Semantics

QLF

Figure 1: The analysis steps of the S-CLE

2. Verb phrase syntax Since the Swedish grammar was developed from an original one for English, it is important to notice that the main differences between English and Swedish word-order stem from the strongly verb-second nature of Swedish: formation of both YN- and WH-questions is by simple inversion of the subject and verb, without the introduction of an auxiliary. For example: Han sŒg Maria. SŒg han Maria? Vem sŒg han?

ÒHe saw Mary.Ó ÒDid he see Mary?Ó ÒWho did he see?Ó

In this and the following section we will see that this difference of main verb syntax can be factored away in the QLF-based semantics, while some auxiliary verb cases still causes some problems. We will however not go into full detail on Swedish verb-phrases in general, they are treated in length by other authors, for example Andersson (1977) and Tjekalina (1991).

A main idea in the unification-based grammar theory used in the CLE is that large quantities of information that has traditionally resided in the grammar rules should be incorporated into the lexicon. For verbs, specifically, removing the information regarding subcategorization schemes (i.e., the number and type of verbal complements, such as objects, particles, etc.) from the grammar enables us to treat both English and Swedish verb-phrases of different kinds in a simple and uniform way, thus: VP --> V ( subcat = Complements ) + Complements which is to be interpreted so that ÔsubcatÕ is a feature of the verb and has the value ÔComplementsÕ, which in turn is unified with the rest of the verb-phrase. The value of ÔsubcatÕ is specified for a particular verb in its lexical entry and can of course be empty (for intransitives, etc.). For example the lexical entry for a transitive particle verb as tycka om nŒgon, ÔÔlike somebodyÕÕ, becomes: V ( subcat = P NP ) Our current Swedish grammar treats 48 different main verb complement patterns plus copulas and auxiliaries. Without claiming this to be the absolute number of Swedish verb types in any sense, it is easily understandable that without the strategy outlined above, we would have been forced to state specific instances of the verb-phrase formation rule for a vast number of cases. Thus the above rule actually is a rule schema, replacing a multitude of verbphrase formation rules which could have been written explicitly. Writing a specific rule for each verb type might not be a completely unfeasible task, but both cumbersome and inelegant. Instead all these verb types can be treated with the single verb-phrase rule, which in fact will be applied (at least) once for the formation of all (complete) sentences. Two things are to be noted about the above treatment: firstly, the ÔÔcatch-allÕÕ verb-phrase formation rule only treats obligatory verbal complements, additional rules are still necessary for optional verb-phrases modifiers, that is certain adverbs, negation, and prepositional phrases, as well as for verb-phrase conjugation. Secondly, some of the gain of having a simplified grammar is of course lost in that the lexicon gets more complicated. In particular this means that verbs that can take several kinds of complements will have to be specified as in effect a number of different verbs in the lexicon.

The main claim of this section is that having only one single verb-phrase formation rule is something worth aiming for. This is defendable if the one rule both saves time and helps avoiding redundancy in the grammar, which is the case if the different verb-types covered are many and diversified. We will thus exemplify the above discussion with some of the different types of verbs actually covered by the S-CLE grammar. In general terms, verbs can be classified by the type of complements they take, thus we have, among others, the verb types shown in Table 1 (and other combinations of these types). Verb type

Example

Impersonals regular impersonal impersonal reflexive

det regnar det ordnar sig

Intransitives regular intransitive intransitive reflexive verb intransitive particle verb intransitive particle reflexive verb intransitive reflexive particle verb intransitive verb, special PP complement intransitive reflexive verb, special PP complement intransitive particle verb, special PP complement intransitive particle reflexive verb, special PP complement

snarka harkla sig falla ned vika ut sig gšra sig till lita pŒ nŒgon gifta sig med nŒgon rŒka ut fšr nŒgot klŠ ut sig till nŒgot

Transitives regular transitive verb which passivises transitive reflexive verb transitive particle verb

gilla nŒgon nŠrma sig nŒgon tycka om nŒgon

Ditransitives regular ditransitive

ge nŒgon nŒgot, ge nŒgot till nŒgon

Propositionals Ð the subject of the proposition is not (necessarily) the subject of the verb. regular propositional pŒpeka att nŒgon gjort nŒgot transitive particle propositional upplysa nŒgon om att nŒgon gjort nŒgot WH-propositional undra vad nŒgon gjort reflexive particle WH-propositional fšrhšra sig om vad nŒgon gjort Infinitivals Ð the subject of the lower VP is the subject (or object) of the verb. bare subject infinitive fšrsška gšra nŒgot bare object infinitive lŒta nŒgon gšra nŒgot ÒattÓ subject infinitive avse att gšra nŒgot Auxiliaries tense modal

har gjort nŒgot kan gšra nŒgot

Copulas verb phrase, past participle noun phrase adjectival phrase prepositional phrase

Šr sŒld Šr en bil Šr ršd Šr pŒ vŠgen

Table 1: Examples of different verb-types in Swedish

As pointed out above, the complements (subcategorizational pattern) of a particular verb is defined in its lexical entry. The lexical entries are quite straight-forward: the one for a transitive particle verb has already been shown, some other relevant examples are the entries for a reflexive particle WHpropositional as fšrhšra sig om vad nŒgon gjort, ÔÔinquire (reflexive) about what somebody has doneÕÕ and for a modal auxiliary as kan gšra nŒgot, ÔÔbe able to (can) do somethingÕÕ, which are, respectively V ( subcat = NP ( reflexive = y ) P S ( type = q ) ) V ( subcat = VP ) 3. Semantics Looking at the semantic side the situation gets a bit more complicated; while main verbs still can be treated easily by a rule parallel to the single syntactic one, care has to be taken while treating auxiliaries. The treatment in the CLE follows the strategy suggested by Alshawi & Crouch (1992): information about tense, modality, etc., is packaged declaratively in the compositional semantics. This information could then be unpacked later on to determine the implicit points in time, etc., not shown in the surface form of the sentence. For a translation system like the SLT, this is not necessarily important: information about tense etc can in many cases be translated as it is to the other language. Part of the treatment here of Swedish verb semantics was guided by the fact that it indeed was going to be used in a translation system, so even though auxiliaries behave quite differently in English and Swedish, the aim was for the semantic representations to be as close to each other as possible (the implementational aspects of the work are further described in GambŠck, 1993). The main verb case simply adds semantic information (within angle brackets) to the syntactic verb-phrase formation rule and becomes: VP ÜVÝ --> V ÜVÝ ( subcat = Complements ) + Complements where the V by unification percolates the daughter verbÕs semantic interpretation up to become the interpretation of the entire verb-phrase, and the feature-value pair Ôsubcat = ComplementsÕ still functions as in the syntactic rule described previously.

The semantic interpretations of the complements are simply unified into the verbÕs semantics in the lexicon. The lexical entry for the transitive particle verb example above (tycka om nŒgon) can schematically be viewed as V Ülx.E(x,NP)Ý ( subcat = P NP ÜNPÝ ) where E represents the interpretation of the event as such, NP the semantics of the object, and the lambda-abstraction is over the semantics of the (sought) subject, x . So a transitive verb is viewed as an event which is a two-place predicate over its subject and its object. The actual treatment of what a specific verbÕs semantic interpretation should be in the lexicon is in fact not that important here Ð most of what is found between the angle brackets could have been replaced with something completely different without affecting the general treatment Ð the important part is how the semantic information is passed up from daughter nodes to mother nodes (i.e., from left-side objects to right-hand side objects in the phrase structure rules). Is should anyhow be noted that the example shown is not the way the main verbs are interpreted in the S-CLE lexicon, even though in a sense it is a somewhat simplified version of the actual implementation. Tense auxiliaries Auxiliaries that change the tense of the verb-phrase (e.g., to past as hade, ÒhadÓ, or future as ska, ÒshallÓ) must be treated separately from the main-verb case above. For main verbs, the tense information of the verb-phrase is the same as the one of the daughter verb and is simply unified up together with the other semantic information. In the auxiliary case, the semantic interpretation of the mother verb-phrase should still be the one of the daughter verb-phrase, but the tense should be taken from the auxiliary. We introduce a parameter t which is having as its purpose to package the tense information but which should not affect the (other) semantic information. Adding this parameter in the form of t-subscripts to the semantic verb-phrase formation rule, we get: VP1 ÜVt1Ý --> V ܯt1Ý ( subcat = VP2 ) + VP2 ÜVt0Ý

with the ¯ indicating that the auxiliary perform no other semantic function besides carrying the tense information (in t1 ). The fact that the verb on the right-hand side is an auxiliary is here indicated by it subcategorizing (explicitly) for a verb-phrase, a verb-phrase holding the true semantic interpretation of the entire (mother) verb-phrase. The value t0 for the tense of the daughter verb-phrase indicates that its tense will not influence that of the mother; however, it may, but this should be treated in the lexical entry for the auxiliary. Normally, this will not be the case, though, and an example of a lexical entry for a tense auxiliary is the one for hade, ÒhadÓ, changing the tense to past: V ÜVPt=pastÝ ( subcat = VP ÜVPt0Ý ) Modal auxiliaries Modal auxiliaries complicate the picture somewhat: we need to treat two cases, one for finite and one for non-finite (i.e., infinite plus supine) verb forms, the difference being that the former can modify other modals as in a sentence like Jag skulle vilja kunna flyga.

ÒI would like to be able to flyÓ (lit. ÒI should want could flyÓ)

Note that this type of modal-modal modification not is possible many languages (it cannot occur in English, for example). In examples like this one (where at least the skulle vilja construction is very common), finite modals behave quite a bit like tense auxiliaries; they do not affect the semantic content as such, but rather a modal parameter m that parallels the tense parameter t introduced in rule for tense auxiliaries above. The version of the verb-phrase formation rule for finite modals is thus: m1

VP1 ÜVt1Ý --> m1

V ܯt1Ý ( subcat = VP2 ) + m0

VP2 ÜVt0Ý Again, the tense and modal information of the daughter verb-phrase may actually influence the mother, but only indirectly through the lexical entry of the modal auxiliary in question. Normally, however, this will not be the case, and a typical lexical entry for a finite modal auxiliary is the one for skulle, ÒshouldÓ, the imperfect (finite) form of ska :

m=ska

V ÜVPt0

m0

Ý ( subcat = VP ÜVP t0Ý )

Non-finite (infinite or supine) modals on the other hand behave just like ordinary verbs in the effect they have on the semantic interpretation proper. They could thus be treated with the same rule instance as the main verbs, so the final version of the verb-phrase formation rule becomes: m

VP ÜVtÝ --> m

V ÜVtÝ ( subcat = Complements ) + Complements Completing the skulle vilja example, the lexical entry for a non-finite modal auxiliary as vilja, ÒwantÓ would show that the modal and tense parameters of the daughter verb-phrase not necessarily carries over to the auxiliary: m

m0

V Ülx.E(x,VP)tÝ ( subcat = VP ÜVP t0Ý ) 4. Negation The discussion above did not cover negation, a construction which in Swedish is expressed with the particle inte (ÔÔnotÕÕ) placed after the main verb in a main clause, but before it in a subordinate clause, thus: Han sŒg inte Maria. ÒHe did not see Mary.Ó Han sade att han inte sŒg Maria. ÒHe said that he did not see Mary.Ó In the English version of the CLE grammar negation is treated semantically as an operator on the sentence structure which at the syntactic level pre-modifies a verb-phrase forming a new verb-phrase, the rule thus being schematically: VP --> not + VP In Swedish such a treatment does not suffice; instead of modifying verbphrases, negation is taken as modifying the verb itself in the syntax. Since whether the modification is pre- or post- depends on the type of clause, the clause type is here indicated by a feature called subordinate.

Three rules for verbs are needed, the first two treating the main clause and subordinate clause cases shown above, and the third treating a special case of main clause negation with a pronoun as object, where the pronoun optionally may occur in front of the negation, Han sŒg mig inte (ÒHe did not see meÓ): V ( subordinate = n ) --> V + inte V ( subordinate = y ) --> inte + V V ( subordinate = n , subcat=Rest) --> V ( subcat=Pro+Rest) + Pro + inte At the semantic level, treating negation as an operator causes some problems. Firstly, a number of other common adverbials (so-called ÔÔmobile adverbsÕÕ), including ofta (ÔÔoftenÕÕ), alltid (ÔÔalwaysÕÕ) and troligen (ÔÔprobablyÕÕ) behave in the same way as inte. Of course, all mobile adverbs ought to be treated in the same way, but introducing specific logical-form operators for all of them would hardly be feasible. Thus inte is actually the only mobile adverb treated by the present version of the Swedish grammar. Secondly, in the verb-phrase formation rule for modals the auxiliary semantics will normally be disregarded, but for non-lexical verbs (i.e., those which have been modified by negation) the semantics will include the modifying operator which must be passed up to the verb-phrase formed. Thus the final version of the non-modal auxiliary and finite modal verb-phrase formation rule as implemented is: m1

VP1 Ü(Aux,V)t1Ý --> m1

V ÜAuxt1Ý ( subcat = VP2 ) + m0

VP2 ÜVt0Ý with ( A u x , V ) being a two-case function for treating auxiliary verb modification, the first case being for no modification (returning the semantics of the daughter verb only) and the second for modified auxiliaries (returning a list with the modifier and the daughter-verb semantics). The operator-treatment of negation, however, causes no problems for the semantics of main verbs and non-finite modals; they same rule instance as given at the end of the previous section can still be used. Acknowledgements: This work was sponsored partially by SICS and partially by the Swedish Telecom, Network Division as part of the ÒSpoken Language TranslationÓ project. I would like to thank my colleagues at SICS Ivan Bretan and Jussi Karlgren for valuable comments.

5. Conclusions and future work The paper has outlined a uniform treatment of Swedish verb syntax and semantics mainly based a rule schema like verb-phrase formation rule extended with specific parameters for tense, mood, etc. Formally, these parameters represent situation-dependent properties that must be applied to the semantic formula in order to give a complete interpretation. Obtaining this full interpretation is, however, not always necessary, and the paper only discussed how the first stages of syntactic and (compositional) semantic interpretation would treat the parameters. Full interpretation is left to later-stage analysis. On the implementation side of the above treatment, it should be noted that even though the rule schema verb-phrase formation rule allows for the treatment of over fifty different verb-types with one single rule, the actual implementation does not treat copulas. These could however easily be covered by the general treatment Ð not doing so is more a practical (efficiency) question than a theoretical one. Theoretically, there is also no reason for not treating both main verbs and auxiliaries with exactly one single verb-phrase formation rule. However, in practice the rule has been split into two cases, one for main verbs and non-finite modals and one for non-modal auxiliaries and finite modals. Merging these two semantic rules into one is also not feasible in the current implementation; this should not be inherent in the theory as such, though, and reaching a completely uniform treatment should not be far off. References H. Alshawi (ed.), 1992. THE CORE LANGUAGE ENGINE. The MIT Press, Cambridge, Mass. H. Alshawi and R. Crouch, 1992. Monotonic Semantic Interpretation. THE 29TH ANNUAL MEETING OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, Newark, Delaware. M-S. AgnŠs, H. Alshawi, I. Bretan, D. Carter, K. Ceder, M. Collins, R. Crouch, V. Digalakis, B. Ekholm, B. GambŠck, J. Kaja, J. Karlgren, B. Lyberg, P. Price, S. Pulman, M. Rayner, C. Samuelsson and T. Svensson, 1991. Spoken Language Translator: First-Year Report. SICS Research Report, R94003, Stockholm. E. Andersson, 1977. VERBFRASENS STRUKTUR I SVENSKAN: EN STUDIE I ASPEKT, TEMPUS, TIDSADVERBIAL OCH SEMANTISK R€CKVIDD. PhD Thesis, •bo Akademi. B. GambŠck, 1993. On Implementing Swedish Tense and Aspect. THE 9TH SCANDINAVIAN CONFERENCE ON COMPUTATIONAL LINGUISTICS, Stockholm University. B. GambŠck and M. Rayner, 1992. The Swedish Core Language Engine. THE 3RD NORDIC CONFERENCE ON TEXT COMPREHENSION IN MAN AND MACHINE, Linkšping University. E. Tjekalina, 1991. Om kategorierna modus och tempus i nutida svenska. SPR•K & STIL, Volume 1, pp 139-155.