Acquisition of a Lexicon from Semantic Representations of Sentences

Cynthia A. ThompsonACL 1995 · DOI

Introduction

A central challenge in natural language processing (NLP) is lexical acquisition: the process by which a system learns the mapping between surface-level words and their underlying semantic meanings. Historically, many systems relied on hand-coded lexicons or required significant prior knowledge of syntax to function. In this paper, Cynthia Thompson introduces WOLFIE (WOrd Learning From Interpreted Examples), a system designed to acquire a lexicon automatically by pairing sentences with their deep semantic representations. By identifying the "common denominator" in tree-based meaning structures across various contexts, WOLFIE mimics the way a child might learn language by associating heard sentences with observed scenes.

The Challenge of Lexical Acquisition

In the mid-90s, most NLP systems were bottlenecked by the need for manual dictionary creation. While some systems attempted to bootstrap word learning, they often required complex grammatical rules or functioned only on highly artificial, simplified data. WOLFIE addresses this by learning from "interpreted examples"—sentences paired with formal representations of their meaning. The goal is to derive a mapping between words and conceptual definitions (represented as trees) without requiring any prior knowledge of nouns, verbs, or grammar.

The WOLFIE Algorithm: Learning by Intersection

The core of WOLFIE is an algorithm that utilizes Tree Least General Generalizations (TLGGs) combined with statistical analysis and a greedy search strategy. The process follows four primary steps:

  1. Candidate Generation: For every word in the training set, WOLFIE examines all sentence-meaning trees in which that word appears. It performs TLGGs—finding the largest common subtrees—between pairs of these representations to generate potential candidate meanings for the word.
  2. Greedy Hill-Climbing: The system employs a greedy search to identify the most probable word-meaning pairs. A candidate meaning is prioritized if it consistently overlaps with a high percentage of the sentence representations where the word occurs.
  3. Iterative Subtraction: To ensure precision, once a word-meaning pair is learned, that specific meaning is "subtracted" from the sentence representations. This prevents the system from mistakenly assigning the same semantic component to multiple words within the same sentence (e.g., once "apple" is mapped to its subtree, that subtree is removed so it is not accidentally associated with the verb "ate").
  4. Handling Ambiguity: WOLFIE is designed to be flexible, explicitly allowing for homonymy (one word having multiple meanings) and synonymy (multiple words sharing a single meaning).

Experimental Setup and Datasets

The system was evaluated using a corpus based on the work of McClelland and Kawamoto (1986), consisting of 1,475 sentence/case-structure pairs. These pairs were generated from 19 sentence templates. However, the authors modified the original case structures to use Conceptual Dependency (CD)—a more complex, tree-based semantic representation developed by Roger Schank. This shift to CD representations provided a more rigorous test of WOLFIE’s ability to handle deep semantic structures.

Key Results and Insights

WOLFIE demonstrated high efficacy in learning lexicons from scratch. The system reached a peak of 94% accuracy in word-meaning pairings after training on 450 examples.

Beyond standalone performance, WOLFIE proved to be a valuable component for broader NLP tasks. When WOLFIE’s learned lexicon was fed into the CHILL parser—a system that previously struggled to map sentences to deep semantic representations on its own—the parser achieved 55% accuracy on a test set. This demonstrates that WOLFIE can successfully bootstrap more complex parsing systems.

An interesting anomaly was observed in the learning curve: accuracy dipped slightly (to approximately 83%) as the training set grew toward 650 examples. The author attributes this to the greedy search occasionally selecting "incomplete" or overly specific meanings early in the process, which then created a cascading effect of errors in later iterations.

Conclusion and Future Work

WOLFIE represents a significant step toward syntax-light lexical acquisition. By deriving meaning purely from the statistical recurrence of semantic subtrees, it avoids the need for pre-defined grammatical categories. However, several hurdles remain for future development:

  • Morphology: Currently, WOLFIE treats different forms of a word (e.g., "run" vs. "running") as entirely unrelated tokens. Future iterations aim to incorporate morphological analysis to group these variations.
  • Scaling: While the CD-based dataset was complex, it was still template-generated. Testing WOLFIE on noisier, unconstrained real-world text is a necessary next step.
  • Incremental Learning: The current batch-oriented approach requires re-processing the entire dataset to update the lexicon. Moving toward an incremental model would better simulate natural language acquisition.
  • Robustness: The system currently assumes the provided semantic representations are 100% correct. Future research must address how the system handles "noisy" or incorrect training data.

Summary updated

BibTeX
@inproceedings{thompson1995acquisition,
  author = {Cynthia A. Thompson},
  title = {{Acquisition of a Lexicon from Semantic Representations of Sentences}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1995},
  doi = {10.3115/981658.981714},
}
Figure