Ambiguity resolution and the retrieval of idioms: two approaches

Erik-Jan van der Linden, Wessel KraaijCOLING 1990

TL;DR

This paper tackles the challenge of resolving ambiguity in natural language processing when encountering idiomatic expressions. The authors propose that idioms should be treated as holistic lexical units and present two models for their representation and retrieval: one extending a two-level morphology framework and another using a connectionist approach. The connectionist model is highlighted for naturally incorporating the "conventionality principle," where idiomatic meanings are prioritized, leading to more efficient processing.

What Problem Does This Paper Address?

Natural language processing (NLP) systems often struggle with ambiguity, particularly when dealing with idiomatic expressions. An idiom, like "kick the bucket," has a meaning that is not derivable from the literal meanings of its constituent words. When such a phrase appears, an NLP system faces a choice: interpret it literally or as an idiom. This paper addresses the problem of how to efficiently and effectively resolve this ambiguity, arguing that a "conventionality principle" should guide the process, favoring the idiomatic interpretation when applicable. The authors aim to demonstrate how this principle can be naturally integrated into computational models of language processing, moving beyond simply stipulating it. They also explore how idioms should be represented in a lexicon, advocating for their storage as single, indivisible units rather than being decomposed.

Methods

The paper presents two distinct computational models for representing and retrieving idioms:

  • Extension of Two-Level Morphology: This approach adapts the concept of "continuation classes" from two-level morphology. In this framework, lexical entries are stored in a trie structure. The novelty here is extending the idea of continuation classes beyond word boundaries to include references to other lexical entries that form an idiom with the current one. An algorithm is outlined for identifying idioms based on this extended lexicon.
  • Localist Connectionist Model: This model is an extension of existing connectionist approaches to lexical ambiguity. It features multiple levels representing wordforms, syntactic features, and semantic features. Idioms are represented as holistic units within the semantic network. The key innovation is how the "conventionality principle" emerges naturally from the model's architecture. When the constituent words of an idiom are encountered, their activation builds up. However, the idiom unit only becomes fully active once all its constituent lexemes are present in the input. This "best-first" processing, where the idiom unit competes with non-idiomatic interpretations, naturally prioritizes the idiomatic reading.

Datasets

Not discussed in the paper. The paper focuses on theoretical models and algorithmic approaches rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Idioms as Holistic Lexical Units: The paper strongly advocates for treating idioms as single, indivisible lexical entries in the lexicon, rather than attempting to decompose their meaning compositionally. This simplifies processing and aligns with psycholinguistic findings.
  • The Conventionality Principle: This principle suggests that when an idiomatic expression is encountered, the idiomatic meaning should be prioritized. The paper argues that this principle shouldn't be an arbitrary rule but should emerge naturally from the processing model.
  • Connectionist Model's Natural Conventionality: The connectionist model is particularly interesting because the conventionality principle arises organically from its architecture. The idiom unit's activation depends on the presence of all its constituent lexemes, making it a strong competitor against non-idiomatic interpretations once the full phrase is recognized.
  • Efficiency through Best-First Processing: Both models aim to improve NLP efficiency by avoiding the need to process all possible interpretations of a phrase. They focus on identifying promising candidates (idiomatic or non-idiomatic) and prioritizing them, thus reducing computational load on subsequent processing stages.

Key Results

The paper presents two models and an algorithm, but does not report quantitative results from empirical evaluations. The key "results" are conceptual and architectural:

  • The proposed extension to two-level morphology provides a mechanism for representing idioms and an algorithm for their retrieval.
  • The connectionist model demonstrates how the conventionality principle can be naturally implemented, leading to the prioritization of idiomatic readings without explicit stipulation.
  • Both models are designed to contribute to more efficient parsing by resolving ambiguity early.

Limitations and Future Work

  • Computational Cost of Connectionist Models: The authors acknowledge that connectionist models, while elegant, often require parallel hardware for realistic implementation.
  • Integration with Syntactic Modules: The connectionist model, in its current form, is primarily focused on lexical and semantic processing. Future work is planned to integrate it with syntactic modules to account for the precise syntactic structure of idiomatic phrases.
  • Semantic Decomposability: The paper explicitly limits its scope to aspects relevant to a "lexical" approach and does not delve into the issue of semantic decomposability of idioms, though it references other work in this area.

Summary updated

BibTeX
@inproceedings{linden1990ambiguity,
  author = {Erik-Jan van der Linden and Wessel Kraaij},
  title = {{Ambiguity resolution and the retrieval of idioms: two approaches}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1990},
}
Figure