Incremental Dependency Parsing

The paper introduces a dependency-based grammar and the associated parser and focusses on the problem of determinism in parsing and recovery from errors. First, it is shown how dependency-based parsing can be afforded, by taking into account the suggestions coming from other approaches, and the preference criteria for parsing are briefly addressed. Second, the issues of the interconnection between the syntactic analysis and the semantic interpretation in incremental processing are discussed and the adoption of a TMS for the recovery of the processing errors is suggested. T H E B A S I C P A R S I N G A L G O R I T H M The parser has been devised for a system that works on the Italian language. The structure that results from the parsing process is a dependency tree, that exhibits syntactic and semantic information. The dependency structure: The structure combines the traditional view of dependency syntax with the feature terms of the unification based formalisms (Shieber 86): single attributes (like number or tense) appear inside the nodes of the tree, while complex attributes (like grammatical relations) are realized as relations between nodes. The choice of a dependency structure, which is very suitable for free word order languages (Sgall et al. 86), reflects the intuitive idea of a language with few constraints on the order of legal constructions. Actually, the flexibility of a partially configurational language like Italian (that can be considered at an intermediate level between the totally configurational languages like English and the totally inflected free-ordered Slavonic languages) can be accounted for with a relaxation of the strong constraints posed by a constituency grammar (Stock 1989) or by constraining to a certain level a dependency grammar. Cases of topicalization, like un dolce di frutta ha ordinato il maestro a cake with fruits has ordered the teacher and in general all the five permutations of the "basic" (i.e. more likely) SVO structure of the sentence are so common in Italian, that it seems much more economical to express the syntactic knowledge in terms of dependency relations. Every node in the structure is associated with a word in the sentence, in such a way that the relation between two nodes at any level is of a head&modifier type. The whole sentence has a head, namely the verb, and its roles (the subj is included) are its modifiers. Every modifier in turn has a head (a noun, which can be a proper, common or pro-noun, for participants not marked by a preposition, a preposition, or a verb, in case of subordinate sentences not preceded by a conjunction) and further modifiers. Hence the dependency tree gives an immediate representation of the thematic structure of the sentence, thus being very suitable for the semantic interpretation. Such a structure also allows the application of the rules, based on grammatical relations, that govern complex syntactic phenomena, as revealed by the extensive work on Relational Grammar. The dependency grammar is expressed declaratively via two tables, that represent the relations of immediate dominance and linear order for pairs of categories. The constraints on the order between a head and one of its modifiers and between two modifiers of the same head are reflected by the nodes in the dependency structure. The formation of the complex structure that is associated with the nodes is accomplished by means of unification: the basic terms are originated by the lexicon and associated with the nodes. There exist principles that govern the propagation of the features in the dependency tree expressed as analogous conventions to GPSG ones. The incremental parser: In the system, the semantic, as well as the contextual and the anaphoric binding analysis, is interleaved with the syntactic parsing. The analysis is incremental, in the sense that it is carried out in a piecemeal strategy, by taking care of partial results too. In order to accomplish the incremental parsing and to build a dependency representation of the sentence, the linguistic knowledge of the two tables is