The Ups and Downs of Lexical Acquisition

Peter M. Hastings, Steven L. LytinenAAAI 1994

Introduction

A persistent challenge in early Natural Language Processing (NLP) was the "unknown word" problem. Systems were often fragile, failing entirely when encountering a lexeme not present in their hand-coded dictionaries. The Camille system (Contextual Acquisition Mechanism for Incremental Lexeme Learning) addresses this bottleneck through incremental lexical acquisition, leveraging existing domain knowledge to infer the meanings of new words from their linguistic context.

Methodology: The Camille System

Camille operates as an extension of the LINK NLP system, utilizing a unification-based chart parser and a structured inheritance hierarchy of concepts. This hierarchy, organized as an "IS-A" tree, allows the system to represent semantic relationships and constraints.

The inheritance hierarchy for entities, illustrating the relationships between places, humans, and physical targets.

The system's core innovation lies in its asymmetric search strategy, which distinguishes between the acquisition of nouns and verbs based on their distinct logical roles in a sentence.

The Noun/Verb Dichotomy

Camille employs a "conservative" approach for nouns and a "liberal" approach for verbs to navigate the concept hierarchy:

  1. Nouns (Conservative): Nouns typically serve as arguments (the things being acted upon). Camille initializes a new noun with the most general concept consistent with the observed constraints. For instance, if a word appears as the object of "destroy," the system identifies it as a "Physical Target" based on the hierarchy shown in Figure 1. It only descends to more specific nodes (e.g., "Building") if subsequent data necessitates it.
  2. Verbs (Liberal): Verbs act as predicates that impose constraints. Camille adopts a highly falsifiable strategy by choosing the most specific hypothesis possible. If the system guesses a specific action like "Arson" and later observes the verb used with a "Human" object, the hypothesis is invalidated, forcing the system to move up the hierarchy to a more general concept like "Attack."
The action hierarchy, showing how specific verbs like 'Arson' or 'Murder' are constrained by their objects and actors.

Evaluation and Results

The system was evaluated using the MUC-3 and MUC-4 Terrorism corpora, characterized by complex, real-world news reports. To test Camille, known words were removed from the lexicon to see if the system could re-acquire them through context.

  • Ambiguous Nouns: Camille correctly hypothesized 5 out of 9 definitions, achieving 67% precision and accuracy. While recall was lower (44%), the system successfully handled disjunctive definitions for words like "lines."
  • Verbs: In a test of 17 unknown verbs across 50 sentences, the system achieved 41% recall and 47% accuracy. While modest, these results significantly outperformed a random baseline (22% recall, 23% accuracy) in a complex domain.

Discussion and Limitations

The Camille system demonstrates that "falsifiability" is a valuable feature in lexical learning. By making specific, testable guesses about verbs, the system can rapidly refine its knowledge. However, the system currently operates under a "closed world" assumption, meaning it can only map words to pre-existing concepts in its hierarchy rather than creating new ones. Additionally, its inductive nature makes it sensitive to "noisy" or ungrammatical input, which can lead to incorrect generalizations. Future work involves integrating deeper semantics, such as causal scripts, to further enhance the acquisition process.

Summary updated

BibTeX
@inproceedings{hastings1994ups,
  author = {Peter M. Hastings and Steven L. Lytinen},
  title = {{The Ups and Downs of Lexical Acquisition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure