Learning Word Meanings by Instruction

AAAI 1996

TL;DR

This paper introduces a novel method for teaching computers the meaning of new words through natural language paraphrases. By treating word meanings as components in a compositional semantic framework, the system can "invert" semantic combination rules to deduce the meaning of an unknown word when given a sentence and its paraphrase. This approach offers a more intuitive and efficient way to build lexical knowledge for NLP systems compared to manual specification or purely inductive methods.

What Problem Does This Paper Address?

Building robust natural language processing (NLP) systems requires a vast amount of semantic knowledge, including word meanings and how they combine to form sentence meanings. Traditionally, this knowledge has been manually encoded by system designers, a process that is slow, expensive, and difficult to scale. While inductive learning methods exist, they often require massive amounts of data and struggle with the "one-shot" learning observed in humans. This paper tackles the challenge of acquiring lexical semantic knowledge efficiently and naturally, aiming to bridge the gap between human-like learning and current NLP capabilities.

Methods

The core of the paper's contribution lies in its "Learning by Instruction" paradigm, specifically focusing on a natural language dialogue approach. When a system encounters an unknown word, it prompts the user for a paraphrase of the sentence containing the unknown word. The system then leverages a compositional semantics framework, where sentence meaning is a function of word meanings.

The key innovation is the ability to invert semantic combination functions. If a sentence's meaning (S) can be expressed as a combination of known words (K) and an unknown word (U), and a paraphrase's meaning (P) is known, and S and P are equivalent, the system can deduce the meaning of U. This is analogous to solving algebraic equations. For example, if "John is hungry" means John + IS + HUNGRY and "John wants to eat" means John + WANTS + TO + EAT, and these are paraphrases, the system can solve for HUNGRY by manipulating these "equations."

The paper formalizes this using labeled, directed acyclic graphs to represent word meanings and combination functions. It defines a novel inversion operator, (n f x y), which, along with unification, extraction, and subordination operators, allows the system to compute potential meanings for unknown words. The algorithm involves computing a "raw graph," merging nodes, and then pruning the results to generate a set of candidate meanings.

Datasets

The paper does not explicitly mention specific datasets used for training or evaluation. Instead, it uses a sample sentence ("John is hungry") and its paraphrase ("John wants to eat") as a running example to demonstrate the algorithm's functionality. The focus is on the methodology and its theoretical underpinnings rather than empirical validation on large corpora.

Key Insights and Ideas

  • Paraphrase as Instruction: The idea that a human-provided paraphrase can serve as a direct "instruction" for learning an unknown word's meaning is powerful. It leverages human linguistic intuition in a structured way.
  • Inverting Semantic Composition: Treating semantic composition as invertible operations, similar to arithmetic, is a clever way to derive unknown components from known ones. This moves beyond simple pattern matching to a more deductive approach.
  • Graph-Based Semantics: The use of labeled directed acyclic graphs for representing meanings and combination rules provides a flexible and expressive framework for computational semantics.
  • Generalizability: The method is designed to work for all parts of speech and can learn word-concept mappings, lexical decompositions, and syntactic-semantic role mappings, highlighting its broad applicability.

Key Results

The paper demonstrates the algorithm's capability by successfully deriving candidate meanings for the unknown word "hungry" given the sentence "John is hungry" and the paraphrase "John wants to eat." The algorithm generates multiple possible solutions, which the authors suggest can be narrowed down using heuristics. For instance, one of the generated meanings for "hungry" is shown to be an "exact duplicate of the human expert's entry." The paper also explores the inverse problem of learning "wants" when "hungry" is known, generating 26 semantic candidates.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Heuristics for Solution Selection: The algorithm generates multiple candidate meanings, and heuristics are needed to select the most plausible one. The paper briefly discusses two heuristics: favoring more specific meanings and favoring more general ones.
  • Syntactic and Semantic Learning Simultaneously: The current approach assumes a computable syntactic parse even with unknown words. Future work could involve learning syntactic and semantic structures concurrently.
  • Meaning Extension: The system currently focuses on learning entirely new meanings. Extending this to recognize when a known word's meaning should be extended (e.g., "hungry for bananas") is a future goal.
  • Paraphrase Ambiguity: The current algorithm assumes unambiguous paraphrases. Future work needs to address how the system can interact with users to disambiguate potentially ambiguous paraphrases.
  • Formal Analysis: The paper calls for a formal analysis of the algorithm's complexity and completeness, particularly concerning the inversion of unification-based operations.

Summary updated

BibTeX
@inproceedings{1996learning,
  title = {{Learning Word Meanings by Instruction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure