A Syntactic Description of German in a Formalism Designed for Machine Translation

Paul SchmidtCOLING 1988

TL;DR

This paper presents a novel syntactic description of German tailored for machine translation, specifically within the Eurotra project's CAT framework. It introduces a stratificational approach with distinct levels for constituent structure (ECS), relational structure (ERS), and interface structure (IS), mapping phrase structures onto functional structures to handle German's complex word order and unbounded dependencies.

What Problem Does This Paper Address?

Machine translation systems require robust syntactic descriptions of languages to accurately parse and generate sentences. German, with its relatively free word order and phenomena like unbounded dependencies (e.g., long-distance scrambling and wh-movement), poses a significant challenge for traditional syntactic formalisms. Existing approaches often rely on global mechanisms or rigid phrase structure rules that struggle to capture these complexities efficiently. This paper aims to provide a functional, stratificational syntactic framework for German that can be integrated into a machine translation system, specifically addressing the need to map surface-level phrase structures to deeper functional representations.

Methods

The paper introduces the CAT (Computational Analysis of Translation) framework, which is built upon a stratificational description of language with three levels: ECS (Eurotra-Constituent-Structure), ERS (Eurotra-Relational-Structure), and IS (Interface Structure). The core of the syntactic description relies on two types of rules:

  • b-rules (structure building rules): These rules define both constituent structures (ECS) and functional structures (ERS), essentially mapping phrase structures onto functional structures.
  • Feature-rules and killer-filters: These rules operate on structures built by b-rules to express generalizations over attributes, enforce well-formedness conditions, and handle feature percolation.

A key aspect is the functional approach, where phrase structures are mapped onto functional structures, contrasting with configurational approaches that rely on universal phrase structure schemas like X-bar. The paper also details how unbounded movement phenomena are handled locally through a series of specialized translation rules (t-rules) and feature checking mechanisms, avoiding global mechanisms.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Stratificational Description: The Eurotra framework's three-level system (ECS, ERS, IS) provides a structured way to represent linguistic information, moving from surface constituent order to deeper functional and semantic roles.
  • Functional vs. Configurational: The paper advocates for a functional approach, emphasizing the mapping of phrase structures to functional structures, which is argued to be more suitable for handling complex word order and dependencies than configurational approaches.
  • Local Treatment of Unbounded Dependencies: The CAT framework's lack of global mechanisms necessitates a local treatment of phenomena like wh-movement and scrambling. This is achieved through a series of carefully designed translation rules (t-rules) and feature checking mechanisms.
  • Symmetry Hypothesis for German: The paper adopts the "Symmetry Hypothesis" (or "Doppelkopfanalyse") for German, positing a single underlying structure for all sentence types (matrix and subordinate) and deriving surface variations through movement rules.
  • Feature Percolation and Killer Filters: Feature percolation (e.g., of index, case, number, gender) is crucial for ensuring the correct interpretation of structures, and killer filters are used to enforce completeness and coherence by eliminating ill-formed structures.

Key Results

The paper demonstrates how the proposed framework can generate syntactic structures for German sentences, including those with complex word order and unbounded dependencies. It provides concrete examples of b-rules, t-rules, and feature-checking rules that illustrate the derivation of both constituent and functional structures. The framework is shown to handle phenomena like verb-fronting, topicalization, and wh-movement by mapping them onto functional slots like 'topic' and ensuring correct co-indexing and feature agreement.

Limitations and Future Work

The authors acknowledge that while the framework provides a local treatment for unbounded dependencies, certain phenomena like "pied piping" and multiple wh-movement might require a more powerful mechanism for transfer to semantic representations. The paper suggests that the functional structures generated are a good basis for mapping onto deep syntactic predicate-argument structures, which would then be enriched by semantic information for transfer. Further work would involve refining these transfer mechanisms.

Summary updated

BibTeX
@inproceedings{schmidt1988syntactic,
  author = {Paul Schmidt},
  title = {{A Syntactic Description of German in a Formalism Designed for Machine Translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure