next up previous
Next: Parsing Up: Grammatical Analysis in a Previous: Introduction


The Grammar

The grammar for OVIS2 contains the grammatical knowledge required to analyse a word graph and to determine what the meaning of the utterance corresponding to the word graph is.

Ideally, in order to construct such a grammar one would take a general-purpose computational grammar of Dutch, and adapt it to the current domain and application. Unfortunately, however, we do not know of any computational grammar for Dutch that could easily be adapted to the present task (which, among others, requires processing of spoken language, extensive coverage of locative and temporal expressions, and the construction of fine-grained semantic representations).

The OVIS-grammar is being developed especially for the purposes of this project. On the one hand this has the advantage that the grammar can be tailored to the specific requirements of the present project. On the other hand, we want to adopt general solutions as much as possible, as this increases the chances that the grammar can be used in other domains as well. Thus, in designing the grammar we seek a balance between short-term goals (a grammar which covers utterances typical for the OVIS-domain and is reasonably robust and efficient) and long-term goals (a grammar which covers the major constructions of Dutch in a general way).

The grammar currently covers some of the more common verbal subcategorization types (intransitives, transitives, verbs selecting a PP, and modal and auxiliary verbs), NP-syntax (including pre- and postnominal modification, with the exception of relative clauses), PP-syntax, the distribution of VP-modifiers, various clausal types (declaratives, yes/no and WH-questions, and subordinate clauses), all temporal expressions and locative phrases relevant to the domain, and various typical spoken language constructs.

The Grammar Formalism

From a linguistic perspective, the OVIS-grammar can be characterized as a constraint-based grammar, which makes heavy use of lexical information. The design of the grammar was inspired to a certain extent by Head-driven Phrase Structure Grammar (HPSG) [18].

The OVIS-grammar formalism is essentially equivalent to Definite Clause Grammar (DCG) [17]. The choice for DCG is motivated by the fact that this formalism provides a balance between computational efficiency and linguistic expressiveness, and the fact that it is closely related to constraint-based grammar formalisms, such as HPSG, Categorial Unification Grammar, and Lexical Functional Grammar. Another important reason to choose DCG instead of a more restricted formalism such as context-free grammar, is the fact that DCG allows the kind of integration of syntax and semantics that is standard in constraint-based formalisms such as HPSG.

Grammar rules consist of a context-free skeleton to which feature-constraints are added. The context-free skeleton is important, as it ensures a reasonable level of processing efficiency and facilitates experimentation with different parsing techniques.

The central formal operation in constraint-based grammar formalisms is unification of (typed or untyped) feature-structures [19]. The OVIS-formalism employs typed feature-structures in the definition of rules as well as lexical entries. During the construction of the parser, feature-structures are translated into Prolog terms. Because of this translation step, parsing can make use of Prolog's built-in term-unification, instead of the more expensive feature-unification.

As in HPSG, generalizations about rules are captured by means of principles. For instance, most rules introduce a kind of head-complement or head-modifier structure. Both structures inherit from the head-feature principle (which states that the HEAD-features of the head-daughter and the mother must be unified) and the slash-principle (which governs the percolation of the feature SLASH, used to account for unbounded-dependencies). Figure 2 illustrates part of the rule hierarchy. An arrow from a box A to a box B indicates that A inherits information from B, where inheritance amounts to unification. In the rules shown in figure 2, the underlined daughter is the head of the rule. Note that we assume that DET is the head of NP. The rules V #MATH185# ARG V, V #MATH186# MOD V, and V #MATH187# V MOD are abbreviations for a finite number of rules (ARG can be instantiated as NP, PP or V for instance, the possible instantiations of MOD depend on whether it precedes or follows its verbal head).

Figure 2: Part of the rule hierarchy

Note that the OVIS-grammar does not use the general rule-schemata of HPSG and thus there is no need for linear precedence statements. On the other hand, the requirement that all rules must have a context-free skeleton implies that the number of rules is larger than in most HPSG fragments. As the rule-set is structured as an inheritance network, this does not give rise to unacceptable redundancy.

As is the case in most lexicalist grammar formalisms, subcategorization is handled lexically. Thus, all major categories (verbs, nouns, prepositions, and determiners) have a feature SUBCAT, whose value is the list of complements they subcategorize for. Expressing this information lexically, instead of using more detailed syntactic rules, has the advantage that idiosyncratic subcategorization requirements need not be stated in the rules (such as the restriction that denken (to think), if it selects a PP-complement, requires this complement to be headed by aan (about); or the fact that komen (to come) may combine with the particle aan (the combination of which means to arrive)). Similarly, all constraints having to do with case marking and agreement can be expressed lexically. Finally, the semantic relation between a head and its dependents can be expressed lexically.

Figure 3 illustrates the grammatical analysis for ik wil om ongeveer vier uur vertrekken (I want to leave around four o'clock). The subject ik combines with a verb-initial clause missing an NP (indicated as S#MATH188#INI/NP) to form a root clause. The finite verb wil combines with a sentential constituent missing an NP and a verb (where the verbal gap is indicated as //V). The idea is that wil may appear in this fronted position only if it combines with a sentence in which a verbal gap occurs which is unifiable with wil. This type of analysis resembles the transformational analysis in terms of verb movement and allows us to give a uniform treatment of main and subordinate clauses. Note that gaps are introduced by means of empty elements (i.e. NP/NP is the empty element corresponding with the fronted subject, and V//V is the empty element corresponding with the fronted verb). The temporal NP is analysed using a special grammar module (see below).

Figure 3: ``ik wil om ongeveer vier uur vertrekken''

The Lexicon

Lexicalist grammar formalisms tend to store lots of syntactic information inside the lexicon. To avoid massive reduplication of identical information in the lexicon, the use of inheritance is essential. A fragment of the lexical inheritance hierarchy is shown in figure 4. The lexical entry for a verb such as vertrekken (to leave) is defined in terms of the templates intransitive and iv-sem; intransitive in its turn is defined in terms of a template verb, which defines the properties common to all verbs.

Figure 4: Part of the lexical hierarchy
...\par\put(25,42){\framebox(20,5){\sc verb}}

Lexical rules provide another means for capturing lexical generalizations. While inheritance can be used to express information present in various lexical entries succinctly, lexical rules are used to express relations that are of an implicational nature (if a lexical entry satisfies constraints C, it is valid to infer the existence of a lexical entry C$\scriptstyle \prime$). At the moment, lexical rules are used to encode inflectional morphology and, to a limited extent, to account for unbounded dependencies.

The Grammar of Temporal Expressions

Temporal expressions occur very frequently in the OVIS domain. It is therefore essential that the grammar provides extensive coverage of these expressions. Some examples of temporal expressions of category NP covered by the grammar are presented in below. These expressions can be combined with prepositions such as op (on), om (at), voor (before), na (after) to form a temporal expression of category PP.

a. drie januari January, 3
b. de derde (januari) the third (of January)
c. maandag, maandagmiddag Monday, Monday afternoon
d. overmorgen the day after tomorrow
e. drie uur, drie uur vijftien three o'clock, three fifteen
f. kwart over drie quarter past three
g. vijftien minuten over (drie) fifteen minutes past (three)

At the moment, the grammar of temporal expressions is accounted for in a separate module of the grammar. Thus, the general principles which hold for other parts of the grammar (such as the head-feature principle and the slash principle) need not apply within this module. We believe that such an approach is motivated by the fact that the grammar of temporal expressions contains many idiosyncrasies (such as the fact that drie uur combines a singular noun with a plural numeral, or the fact that vijf over drie combines two numerals and a preposition) which seem to be limited to this type of expression only. Note that this does not imply that no principled account of the syntax of temporal expressions could be given (see [10] for instance). However, as general principles of the grammar do seem to play only a minor role in the grammar of temporal expressions, we are satisfied with providing a few simple, but idiosyncratic, rules and lexical entries to account for these expressions.


The output of the grammatical analysis is a semantic, linguistically motivated and domain-independent, representation of the utterance. This representation is translated into one or more domain-specific updates, which are passed on to the pragmatic interpretation module and dialogue manager (DM) for further processing. Below, we motivate our choice for quasi logical forms (QLF's) as semantic representation language and we discuss how these forms are translated into updates.

The semantic representation language

Predicate logic is often used to represent the meaning of sentences. Due to its long tradition in describing semantics of natural languages it is now a well-established and well-understood technique. Nevertheless, it often seems helpful to extend the predicate calculus with more advanced techniques, such as generalised quantifiers or discourse markers. For OVIS we use the QLF-formalism [1], whose most important feature is that it makes extensive use of underspecification to represent ambiguities.

A property of artificial languages like predicate logic is that they are unambiguous. An ambiguous natural language utterance will therefore correspond with more than one expression in predicate logic, one for each reading of the utterance. The disadvantage of this approach is that for very ambiguous inputs, expensive computations must be carried out to compute all readings. The alternative adopted in the QLF-formalism is to represent ambiguity by means of underspecification and to postpone the computation of all possible readings to the disambiguation component.

A QLF is a description of one or more predicate logical formulas. While building a QLF for an utterance, information is added monotonically [8]. Whenever information is added, the amount of underspecification will decrease and, consequently, the resulting QLF will describe a smaller set of logical formulas. After grammatical processing of an utterance is completed, several knowledge sources (such as an algorithm for generating quantifier scope or for finding referents for anaphoric expressions) may be used to further reduce the amount of underspecification.

In figure 5 we give the (simplified) QLF which the current grammar produces for the example in figure 3. Remember that the grammar manipulates feature-structures, so the QLF is represented as a feature-value matrix. Note that the arguments of the predicates willen and vertrekken are (generalized) quantifiers that are unscoped with respect to each other. Furthermore, we assume that verbs introduce an (existentially quantified) event-variable.

Figure 5: (Simplified) QLF for 'Ik wil om ongeveer vier uur vertrekken'

The QLF in figure 5 represents the domain-independent meaning of the utterance. If we want to reuse the present grammar for other domains, or for extensions of the present domain, it is essential to have representations that are not based on domain-specific assumptions. Representing ambiguity by means of underspecification is particularly important in the present setting, as the system must deal with highly ambiguous input (i.e. word graphs). Furthermore, it turns out to be the case that the computation of the domain-specific meaning of an utterance can be carried out without resolving all the sources of ambiguity in the QLF for that utterance. For instance, the scope of quantifiers seems to be to a large extent irrelevant for the computation of (domain-specific) updates, and thus a representation in which quantifier-scope is undetermined is sufficient for our purposes.

Domain-specific interpretation

The dialogue manager (DM) expects the linguistic module to produce so-called updates representing the domain- and context-specific meaning of the user-utterance. The DM keeps track of the information provided by the user by maintaining an information state or form. This form is a hierarchical structure, in which there are slots for the origin and destination of a connection, for the time at which the user wants to arrive or leave, etc. Updates specify the values of these slots. For example, the QLF in figure 5 is translated into the following update:[=4]

The first part of the update (userwants...clock-hour) is a path specifying the slot that is filled, the second part (4) is the value.

As the user responds to the system's questions, responses will usually not be a complete sentence, but an NP or PP, or a sequence of such constituents. In such cases, the QLF produced by the grammar will be severely underspecified. An update can be constructed, however, by adding information from the question to the semantic representation of the answer. The response in (b), for instance, can be translated into the update in (c) by taking into account the fact that this was a response to the question in (a).

a. Naar welk station wilt u vanuit Groningen reizen?
  To which station do you want to travel from Groningen
b. Amsterdam

next up previous
Next: Parsing Up: Grammatical Analysis in a Previous: Introduction
Noord G.J.M. van