Encoding and Acquiring Meanings for Figurative Phrases

Michael G. Dyer, Uri ZernikACL 1986 · DOI

TL;DR

This paper introduces RINA, a natural language processing system designed to tackle the "final boss" of linguistics: figurative language. By moving beyond word-by-word analysis to a phrasal lexicon, RINA can not only parse complex idioms like "carry the water" or "throw the book at him" but also autonomously learn their meanings when they are first encountered in context.

The Challenge of Figurative Language

Traditional NLP systems often struggle with the sheer idiosyncrasy of human speech. While a computer might easily understand "John ate the apple," it trips over "John kicked the bucket" or "John put his foot down."

The authors identify three specific hurdles:

  1. Phrase Interaction: How do the meanings of two idioms merge in a single sentence (e.g., "He made up his mind to take her to task")?
  2. Phrase Ambiguity: How does a system decide if "he ran into a friend" is a literal collision or a chance meeting?
  3. Phrase Idiosyncrasy: Why is "laying down the law" appropriate in a workplace dispute, while "reading the riot act" implies a specific kind of angry reaction to the past?
Figure 1: Examples of idiomatic variations in a social context (S3-S5).

The core challenge is that figurative phrases are pervasive, yet their meanings are rarely the sum of their individual words.

Methods: The Phrasal Lexicon

The authors propose a Phrasal Lexicon approach, implemented in the RINA system. Instead of just storing words, the lexicon stores "triples":

  • Linguistic Pattern: The syntactic structure (e.g., ?x throw <the book> at ?y).
  • Situation: The context or "pre-conditions" required for the phrase to make sense (e.g., an authority figure punishing a subordinate).
  • Concept: The actual semantic meaning or "result" of the phrase.

The Learning Loop

When RINA encounters an unknown phrase, it follows a "failure-driven" learning path:

  1. Literal Attempt: It first tries to interpret the phrase literally (e.g., "They moved water?").
  2. Contextual Conflict: It realizes the literal meaning doesn't fit the current "goal/plan" context (e.g., why would politicians be literally carrying buckets during a tax reform debate?).
  3. Hypothesis Formation: Using the surrounding context, RINA hypothesizes that "carrying the water" must mean "acting as an agent for a rival's plan."
  4. Refinement: Through further examples, it generalizes the phrase from a specific domain (politics) to an abstract situation (rivalry/agency).

Key Insights and Ideas

  • Situation vs. Concept: A phrase's meaning is separate from the context required to use it. This allows the system to disambiguate phrases that look similar but function differently.
  • Literal Failure as a Signal: The system uses the "nonsense" of a literal interpretation as a trigger to look for a metaphor. If the literal action doesn't support a high-level goal in the story, it's a red flag that the phrase is figurative.
  • Knowledge-Based Learning: Learning is conceptual rather than statistical. To understand "throw the book," the system needs a "Trial Script" (Judge, Defendant, Prosecutor). Once learned, it can abstract that script into a general "Authority Relationship," allowing it to understand a school dean "throwing the book" at a cheating student.

Key Results

The paper qualitatively demonstrates that RINA can:

  • Disambiguate between similar phrases like "put his foot down" and "lay down the law" by checking if the actors in the context have equal or unequal authority.
  • Acquire the meaning of "carry the water" through a simulated dialogue with a user, moving from a literal misunderstanding to a correct conceptual representation of "agency in a rivalry."
  • Generalize meanings from specific scripts (like a courtroom) to abstract social situations (like a boss and employee).

Limitations and Future Work

The authors acknowledge that RINA is not yet "socially mature." Its primary limitations include its inability to handle puns and irony, where literal and figurative meanings overlap. Furthermore, the system faces a "knowledge bottleneck," requiring an immense library of common-sense knowledge about human goals and emotions to function at a human level.

Ultimately, the authors conclude that while a phrasal lexicon increases the size of the dictionary, it simplifies parsing by providing "pre-packaged" chunks of meaning that reduce the need for complex inference during a conversation.

Summary updated

BibTeX
@inproceedings{dyer1986encoding,
  author = {Michael G. Dyer and Uri Zernik},
  title = {{Encoding and Acquiring Meanings for Figurative Phrases}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1986},
  doi = {10.3115/981131.981148},
}
Figure