Building a Large-Scale Knowledge Base for Machine Translation

Kevin Knight, Steve K. LukAAAI 1994

Abstract

In the early 1990s, Knowledge-Based Machine Translation (KBMT) faced a significant "knowledge bottleneck": the requirement for massive, handcrafted world models to handle general-purpose text. This paper details the construction of the PANGLOSS ontology, a 70,000-node knowledge base built using semi-automatic algorithms to merge disparate linguistic resources. By integrating WordNet, the Longman Dictionary of Contemporary English (LDOCE), and other specialized models, the authors demonstrated a scalable path forward for MT systems.

Introduction

While KBMT systems have historically excelled in narrow, technical domains, their expansion into general-purpose translation (e.g., news articles) was long hindered by the difficulty of creating a comprehensive "world model." Such models provide the structured map of objects, processes, and relations necessary for deep semantic understanding. Traditionally, these were built manually—a process that was both slow and expensive.

The PANGLOSS project sought to break this bottleneck by shifting from manual construction to semi-automatic acquisition. By leveraging existing digital resources, the researchers aimed to create a unified ontology that could serve as the semantic backbone for translation between multiple languages.

Methodology and Resource Integration

The construction of the PANGLOSS ontology followed a multi-stage "plan of attack" that combined manual oversight with automated matching algorithms.

The multi-stage merging process for constructing the PANGLOSS ontology.

1. Establishing the Ontology Base (OB)

The process began with the manual merging of the PENMAN Upper Model (a linguistically motivated framework) and ONTOS (a top-level ontology designed for MT). This created a stable, high-level structure of approximately 400 nodes upon which the rest of the knowledge base could be built.

2. Merging WordNet and LDOCE

To populate the ontology with thousands of specific concepts, the authors merged WordNet (a semantic database) with the Longman Dictionary of Contemporary English (LDOCE). Two primary algorithms were employed:

  • Definition Match Algorithm: This algorithm calculated similarity scores between word senses by analyzing shared vocabulary in definitions and examples, as well as synonyms and hypernyms.
  • Hierarchy Match Algorithm: This exploited the taxonomic structures of both databases. It operated on the principle that if parent concepts (hypernyms) match, their respective descendants are also likely to match.

3. Bilingual Integration

To extend the ontology's utility to machine translation, a Spanish lexicon was integrated using the Collins Spanish-English Dictionary. Disambiguation was achieved by matching dictionary subject field codes (e.g., "military" or "medical") to corresponding branches within the ontology.

Key Insights: The Human-in-the-Loop

A central philosophy of the PANGLOSS project was the efficiency of human verification over human generation. The authors found that while algorithms could propose thousands of matches, human experts were far more efficient at quickly confirming or rejecting these proposals than at creating them from scratch.

Furthermore, the project highlighted the value of resource redundancy. By merging WordNet (strong hierarchy) with LDOCE (rich syntactic and subject-area data), the resulting "lattice" became more robust and informative than any single source could be on its own.

Results and Performance

The semi-automatic approach yielded a knowledge base of unprecedented scale for its time, containing roughly 70,000 nodes.

The Hierarchy Match algorithm proved exceptionally reliable, generating over 11,000 noun sense matches with an accuracy rate of 96%. For the Definition Match algorithm, the researchers observed a clear trade-off between the strictness of the confidence threshold and the resulting data coverage.

Trade-off between confidence level, accuracy, and coverage in the Definition Match algorithm.

As shown in the data, setting a high confidence threshold (≥ 0.8) resulted in 90% accuracy but only covered 27% of the potential matches. Lowering the threshold to 0.0 allowed for 100% coverage but reduced accuracy to 75%.

Conclusion and Future Directions

The PANGLOSS project demonstrated that the "knowledge bottleneck" could be bypassed through the intelligent reuse of existing linguistic data. While successful, the authors noted that the system remained sensitive to errors within the source hierarchies (error propagation) and was limited by gaps in the source data (e.g., missing definitions in WordNet). Future work suggested expanding these techniques to other languages like Japanese and exploring the extraction of semantic constraints directly from unstructured text.

Summary updated

BibTeX
@inproceedings{knight1994building,
  author = {Kevin Knight and Steve K. Luk},
  title = {{Building a Large-Scale Knowledge Base for Machine Translation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure