Corpus-Based Knowledge Representation

Alon Y. Halevy, Jayant MadhavanIJCAI 2003

Abstract: Traditional knowledge representation (KR) relies on the manual construction of logically consistent knowledge bases, a process that is often prohibitively expensive and results in brittle systems. This paper proposes a shift toward "Corpus-Based Knowledge Representation," which leverages large, uncoordinated collections of schemas, ontologies, and data fragments. By applying statistical analysis and machine learning to this corpus, systems can automate complex tasks such as schema matching and query answering across heterogeneous sources. Experimental results from the LSD system demonstrate that this approach can achieve high accuracy (75-90%) while avoiding the traditional knowledge acquisition bottleneck.

Introduction

For decades, the Artificial Intelligence community has sought to represent human knowledge in a form that machines can manipulate. The standard approach involves experts painstakingly articulating domain knowledge into formal logic to create a single, coherent Knowledge Base (KB). However, this methodology faces two significant hurdles: the "Knowledge Acquisition Bottleneck" and the challenge of "semantic heterogeneity."

Building traditional KBs is slow and expensive. Furthermore, the resulting systems are often "brittle at the edges," failing when they encounter scenarios not explicitly anticipated by their designers. In the modern era of the Web and large-scale enterprise data, the problem is compounded by semantic heterogeneity—different stakeholders describe the same data using vastly different terms and structures. Merging these disparate, independently designed KBs into a single logical whole is often an intractable task.

The Corpus-Based Paradigm

In their 2003 IJCAI paper, Alon Halevy and Jayant Madhavan propose treating knowledge representation more like Natural Language Processing (NLP) or Information Retrieval (IR). Instead of striving for a single "gold standard" KB, they suggest building a Corpus of knowledge fragments.

This corpus is not a unified logical model but a massive, uncoordinated collection of database schemas, XML DTDs, existing ontologies, instance data, and even historical user queries. Rather than using formal logic to reason over this data, the system computes statistics to identify patterns. Key metrics include:

  • Term Usage: Analyzing how often specific terms (e.g., "GPA") appear as column names versus actual data values.
  • Co-occurrence: Identifying which attributes or concepts tend to appear together across different schemas.
  • Similarity Vectors: Measuring how closely a new schema element resembles known elements within the corpus.

System Architecture

The architecture of a corpus-based KR system is designed to bridge the gap between raw data fragments and high-level applications. At its core, the system relies on a statistical engine that processes the corpus to provide various query interfaces.

Architecture of a corpus-based knowledge representation system.

As shown in the architecture above, the system uses the corpus and its derived statistics to support multiple interface types, such as factual queries and similarity-based queries. These interfaces, in turn, power applications like automated web query answering and schema authoring. This layered approach allows the system to remain flexible and extensible, adding new query types or applications as needed without requiring a total redesign of the underlying knowledge model.

Key Insights and Ideas

The shift to a corpus-based approach introduces several transformative concepts to KR:

  • Patterns over Design: If a corpus is sufficiently large, the statistical patterns within it become more useful for real-world tasks than a perfectly designed, hand-crafted ontology.
  • Multiple Perspectives as Data: Traditional KR views conflicting models as errors to be resolved. Corpus-based KR views them as valuable data points that illustrate the diverse ways a domain can be modeled.
  • Bridging the "Structure Chasm": By providing a map of how terms are used in structured contexts (like databases), a corpus of schemas helps bridge the gap between structured data and unstructured text.
  • Ranked Answers: Unlike the Boolean logic of traditional KR, this paradigm returns a ranked list of plausible answers based on statistical likelihood, mirroring the behavior of modern search engines.

Evaluation and Results

While primarily a position paper, the authors draw on experimental evidence from their LSD (Learning Source Descriptions) system. The LSD system was designed to test schema matching on various Web data sources.

Key findings include:

  • High Accuracy: The system achieved 75-90% accuracy in matching elements between disparate data sources.
  • Complementary Strengths: Corpus-based matching identified correct mappings that traditional heuristic-based methods missed, demonstrating the unique value of the "wisdom of the crowd" inherent in the corpus.
  • Efficiency: By avoiding the laborious construction of manual KBs, the approach becomes feasible for broad, web-scale applications where traditional KR would be cost-prohibitive.

Conclusion and Future Work

Corpus-based KR is not intended as a total replacement for traditional methods. Tasks requiring intricate logical inference—such as medical diagnosis or interpreting complex tax laws—still necessitate finely-tuned, logic-based KBs. However, for the vast majority of data integration and web-scale knowledge tasks, the statistical approach offers a more scalable and robust alternative.

Future challenges include the "cold start problem" of collecting an initial corpus large enough to be effective and determining how to manage "noise" when including domain models from unrelated fields. Despite these challenges, the corpus-based approach represents a significant step toward more flexible and automated knowledge management systems.

Summary updated

BibTeX
@inproceedings{halevy2003corpus,
  author = {Alon Y. Halevy and Jayant Madhavan},
  title = {{Corpus-Based Knowledge Representation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2003},
}
Figure