Acquiring Lexical Knowledge from Text: A Case Study

Paul S. Jacobs, Uri ZernikAAAI 1988

Abstract

In the late 1980s, Natural Language Processing (NLP) faced a significant "lexical gap": systems were limited by hand-coded dictionaries and failed when encountering unknown words. This paper presents a method for automated lexical acquisition by combining two systems—TRUMP and RINA—to "guess" word meanings through morphological analysis, syntactic clues, and contextual reinforcement. Using news text from the Dow Jones News Service, the researchers demonstrate how a system can gradually refine its understanding of complex terms like "merger" without human intervention.

The Challenge of the Lexical Gap

During this era of NLP, building lexicons was a labor-intensive manual process. Systems often lacked the flexibility to adapt to specific domains, such as finance or law, where words carry specialized meanings. The authors sought to overcome this bottleneck by creating a system capable of automated lexical acquisition: learning new words on the fly, refining their definitions over time, and handling the inherent ambiguity of human language.

Methodology: TRUMP and RINA

The proposed approach utilizes two specialized programs within a larger system called SCISOR (System for Conceptual Information Summarization, Organization, and Retrieval):

  1. TRUMP (Text Processing): When an unknown word is encountered, TRUMP generates a Hypothesis Chart. It analyzes morphology (e.g., identifying the "-er" suffix), syntax (determining if the word is a noun or adjective), and semantics to create a set of competing guesses.
  2. RINA (Language Acquisition): RINA processes these hypotheses by examining the surrounding context. For instance, the word "another" serves as a linguistic clue, signaling that the current event is related to a previously known concept in the system's memory.
  3. Gradual Refinement: Rather than committing to a definition immediately, the system retains multiple hypotheses. Subsequent encounters with the word allow the system to prune incorrect guesses and "variabilize" the meaning into a general template (e.g., moving from a specific "merger offer" to a general "Company X's merger").
Morphological hypothesis generation for the word "merger," showing competing interpretations as a noun, comparative adjective, or actor.

Experimental Setup and Datasets

The study utilized a corpus of approximately 40,000 words from the Dow Jones News Service dated January 25, 1988, focusing specifically on corporate takeovers. On this day, 33 "new" words appeared at least seven times, providing sufficient data for the system to attempt acquisition.

Key Insights

  • Contextual Mapping: Linguistic clues like "another" were vital for mapping the unknown word "merger" to the known concept of a "transaction."
  • Morphological Ambiguity: The system initially hypothesized that "merger" could be a comparative adjective (like "larger") or an actor (like "buyer"). It required multiple examples to correctly identify it as a noun representing the transaction itself.
  • Domain Customization: Automatic acquisition allows the system to learn the specific nuances of a field. In the context of takeovers, the system learns that a "merger" is a specific type of company acquisition rather than a general combination.

Results and Categorization

The researchers categorized the 33 new words into three groups based on the difficulty of acquisition:

  • Group A (Easy): Words solvable via simple prefix or root analysis (e.g., "unchanged").
  • Group B (Medium): Words that were syntactically clear but semantically obscure (e.g., "debentures").
  • Group C (Hard): Words requiring heavy reliance on surrounding context to differentiate meanings (e.g., "crude," "dividend").
The categorization of 33 new words into three groups (A, B, and C) based on the complexity of their automated acquisition.

Limitations and Future Work

The system's learning is sensitive to the order of examples; a confusing initial sentence can lead the system down an incorrect path. Furthermore, because the system learns from a specific domain, its definitions can be "skewed" (e.g., not understanding "merging" in a traffic context). The authors concluded that future work should focus on more robust morphological databases and generalized learning algorithms capable of handling a wider variety of linguistic structures.

Summary updated

BibTeX
@inproceedings{jacobs1988acquiring,
  author = {Paul S. Jacobs and Uri Zernik},
  title = {{Acquiring Lexical Knowledge from Text: A Case Study}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure