A Quantitative Model of Word Order and Movement in English, Dutch and German Complement Constructions

Karin Harbusch, Gerard KempenCOLING 2002

TL;DR

This paper introduces "Performance Grammar" (PG), a novel quantitative model that uniformly explains complex word order and movement phenomena across English, Dutch, and German complement constructions. By leveraging typed feature unification and a system of "topologies" with adjustable parameters, PG offers a parsimonious account of linguistic variation, reducing it to simple quantitative differences.

What Problem Does This Paper Address?

Linguistic theory has long grappled with explaining the diverse ways words are ordered and how phrases move within sentences, especially across different languages. Phenomena like Wh-extraction, clause union, and extraposition in languages like English, Dutch, and German present a seemingly heterogeneous set of challenges. Traditional approaches often require distinct mechanisms for each phenomenon or language. This paper aims to bridge this gap by proposing a single, unified, and quantitative model that can account for these variations through a consistent underlying framework. The core problem is to find a formal system that is both descriptive (capturing the data) and explanatory (providing a principled account of the underlying mechanisms) for cross-linguistic word order and movement.

Methods

The paper's core contribution lies in its Performance Grammar (PG) formalism. PG models sentence structure using hierarchical structures composed of lexical frames. These frames are built from segments and have a three-tiered structure: a phrasal root, functional nodes (like SUBJect, HeaD), and phrasal "feet" dominated by functional nodes.

The key innovation for word order and movement is the concept of topologies. Each phrasal root node in a lexical frame is associated with a feature matrix specifying a one-dimensional array of slots representing the linear order. For S-nodes (clauses), these topologies are proposed to have nine slots, categorized into Forefield, Midfield, and Endfield.

Movement and linearization are handled through lateral topology sharing. When a complement clause's topology shares slots with its governing clause's topology, constituents in those shared slots can appear to have "moved" to the higher clause level. This sharing is controlled by quantitative parameters (LS for left-peripheral sharing, RS for right-peripheral sharing) that are set in the lexical entries of verbs. The paper uses typed feature unification as the mechanism for combining these structures.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and uses linguistic examples rather than formal datasets for training or evaluation.

Key Insights and Ideas

  • Unified Treatment of Heterogeneous Phenomena: The most significant insight is that seemingly disparate word order and movement phenomena (Wh-extraction, clause union, extraposition, verb clustering) across English, Dutch, and German can be explained by a single, unified framework.
  • Quantitative Parameters Drive Variation: Linguistic differences are not attributed to fundamentally different grammatical mechanisms but to quantitative variations in parameters like the size of shared topological areas (LS and RS). This offers a more parsimonious explanation.
  • Topologies as Linearization Mechanism: The concept of "topologies" provides a concrete mechanism for mapping hierarchical structures onto linear word order, with slots acting as "landing sites" for constituents.
  • Lateral Topology Sharing for Movement: The paper elegantly explains movement (promotion) as a consequence of shared topological slots between clauses, where constituents in shared slots are processed at the higher clause level.
  • Parsimony and Computational Efficiency: The authors argue that their approach is theoretically parsimonious and computationally efficient, avoiding the need for separate mechanisms for different phenomena, which is common in other frameworks like HPSG.

Key Results

The paper demonstrates how the PG model, through the settings of LS and RS parameters, can account for:

  • English: Phenomena like Wh-fronting (e.g., "Who do I have to call?") and VP Extraposition.
  • Dutch and German: Similar phenomena, including scrambling and verb clustering, highlighting how differences in parameter settings (e.g., LS=0 for Dutch interrogatives vs. LS=1 for German interrogatives) lead to distinct surface orders.
  • Cross-linguistic Similarities: The model shows that the underlying mechanisms for complement clause linearization are remarkably similar across the three languages, with variations arising from specific parameter values.

Specific quantitative results are not presented in terms of metrics like accuracy or speedup, as the paper is primarily a theoretical proposal. However, the "results" are the successful modeling of numerous linguistic examples.

Limitations and Future Work

  • Generalizability to Other Languages: The authors acknowledge that future work is needed to determine if the PG approach can generalize to other languages beyond English, Dutch, and German.
  • Fine-grained Discussion: A more detailed discussion of the approach and its psycholinguistic motivation has been provided elsewhere (Kempen & Harbusch, in press; forthcoming).
  • Computational Implementation: While a PG sentence generator for Dutch exists, further exploration of computational models for parsing based on PG is suggested.
  • Agreement Features: The paper mentions agreement features but states they are not discussed here, implying a potential area for future research.
  • Island Effects: The model successfully predicts "island effects" (e.g., in sentence (4)), but the precise mechanisms and constraints governing these effects could be further elaborated.

Summary updated

BibTeX
@inproceedings{harbusch2002quantitative,
  author = {Karin Harbusch and Gerard Kempen},
  title = {{A Quantitative Model of Word Order and Movement in English, Dutch and German Complement Constructions}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure