A Practical Classification of Multiword Expressions

Rados aw MoszczynskiACL 2007

Abstract

Multiword Expressions (MWEs) present a significant challenge for Natural Language Processing (NLP) due to their idiosyncratic nature. This paper proposes a functional, two-tier taxonomy for MWEs based on their optimal placement within an NLP pipeline. By distinguishing between expressions that violate basic grammatical rules and those that follow grammar but possess non-compositional meanings, the proposed framework aims to improve parsing coverage and efficiency, particularly in inflection-heavy languages like Polish.

Introduction

Multiword Expressions—ranging from idioms like "kick the bucket" to fixed phrases like "by and large"—are notoriously difficult for computational systems. These units are often lexically variable yet syntactically constrained, frequently violating the standard rules of grammar that parsers rely on. Historically, researchers have struggled to find a balance between over-simplifying these expressions and over-engineering the rules required to handle them.

This paper addresses the need for a practical, computationally efficient way to classify MWEs. The author argues that the primary failure of previous formalisms lies in a "one-size-fits-all" approach. Simple, frozen expressions are often burdened with complex syntactic descriptions, while flexible, complex expressions are forced into rigid templates that cannot accommodate the free word order characteristic of languages like Polish.

A Functional Taxonomy of MWEs

Rather than classifying MWEs by linguistic category (e.g., noun phrase vs. verb phrase), the author proposes a taxonomy based on when an NLP system should process them. This divides MWEs into two functional groups:

1. Morphosyntactically Idiosyncratic Expressions

These units should be handled during the preprocessing stage, before the syntactic parser begins its work. Treating these as single units (similar to named entities) prevents the parser from attempting to apply standard rules to non-standard structures. This category includes:

  • Morphological Idiosyncrasy: Phrases containing "bound words" that only exist within that specific context or take on irregular forms.
  • Syntactic Idiosyncrasy: Phrases that break standard grammar rules (e.g., "Get your hands off!").
  • Complex Adverbials: Phrases that appear to be standard words but can mislead a parser into making incorrect connections if not treated as a single unit.

2. Semantically Idiosyncratic Expressions

These expressions follow regular grammatical patterns but possess non-compositional meanings (idioms). Because they allow for significant variation in word order and internal modification—especially in inflectional languages—they should be recognized simultaneously with syntactic analysis. The author suggests that these are best handled via valence dictionaries during the parsing process.

The Challenge of Inflectional Languages

A key insight of the paper is that highly inflectional languages with free word order, such as Polish, act as a "stress test" for MWE theories. Systems designed primarily for English, which has a relatively rigid word order, often fail when applied to Polish. The author uses a study of 300 multiword units to demonstrate that existing formalisms struggle with the "unification" required to enforce grammatical agreement between words in a phrase.

Critique of Existing Formalisms

The paper evaluates two prominent formalisms:

  • IDAREX: Based on regular expressions, IDAREX is found insufficient for Polish because it lacks the ability to handle grammatical unification.
  • Phrase Manager: Based on syntactic trees, this approach leads to a "proliferation of classes." As more MWEs are added, the system becomes unmanageable due to the sheer number of specific syntactic rules required.

Proposed Pipeline and Results

The author concludes that the most viable path for high-coverage NLP is a hybrid pipeline. This involves a "named entity-style" preprocessor (utilizing tools like lxtransduce) to handle morphosyntactic idiosyncrasies, followed by a grammar-aware parser that identifies semantic idiosyncrasies through valence dictionaries.

While the paper is primarily theoretical and does not provide quantitative F1 scores, its qualitative results highlight a path toward avoiding over-engineering. By treating simple, frozen interjections as preprocessed units rather than complex syntactic trees, researchers can significantly reduce computational overhead.

Limitations and Future Work

The study was specifically limited to Polish and English examples. While the author believes the findings are broadly applicable, further research is needed to test the taxonomy across a wider variety of language families. Future work will focus on the implementation of the suggested syntactic preprocessor and the formalization of these units into a comprehensive valence dictionary for verbs.

Summary updated

BibTeX
@inproceedings{moszczynski2007practical,
  author = {Rados aw Moszczynski},
  title = {{A Practical Classification of Multiword Expressions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2007},
}
Figure