Ontology-Based Information Extraction with a Cognitive Agent

Peter Lindes, Deryle W. Lonsdale, David W. EmbleyAAAI 2015 · DOI

Abstract

OntoSoar is a machine reading system that leverages the Soar cognitive architecture to extract structured genealogical data from idiosyncratic historical texts. By integrating link grammar parsing with cognitive semantics, it transforms noisy OCR-processed records into populated ontologies. This paper details the architecture of OntoSoar and evaluates its performance on large-scale historical datasets, demonstrating high precision in extracting complex familial facts.

Introduction

Historical family records represent a significant untapped data source, often "trapped" in scanned PDF formats. Extracting information from these documents is challenging due to non-standard syntax (e.g., "b. 1817, d. 1865"), OCR noise, and the prevalence of implicit information. Standard Information Extraction (IE) tools often fail on such "sub-languages." OntoSoar addresses these hurdles by treating IE as a reasoning task within a cognitive architecture, enabling deep reading and inference.

System Architecture

The OntoSoar pipeline is built within the Soar cognitive architecture, moving from raw text to structured ontology mapping through several specialized stages.

The OntoSoar system architecture, showing the pipeline from PDF text to a populated user ontology.

The process begins with Preprocessing & Tokenization, where domain-specific abbreviations are expanded and OCR errors are corrected. A Link Grammar (LG) Parser then identifies pairwise word associations, a method better suited for telegraphic genealogical styles than traditional constituent parsing.

The Meaning Builder, based on Embodied Construction Grammar (ECG), uses Soar productions to map these linguistic structures to conceptual "meaning schemas." Subsequently, the Conceptual Semantic Analyzer performs higher-level reasoning, such as pronoun resolution and gap-filling via inference (e.g., inferring a spouse from a marriage mention). Finally, the Mapper projects these internal representations onto an Object-oriented Systems Modeling (OSM) ontology.

Evaluation and Results

OntoSoar was evaluated using gold-standard manually annotated pages, randomized samples from a vast private repository, and a full 830-page volume, The Ely Ancestry.

Performance metrics for OntoSoar across different genealogical categories, showing high precision but lower recall for complex structures.

The system achieved a high overall precision of 92.9%, with near-perfect accuracy in identifying death events and persons. However, recall was lower at 67.5%, primarily due to difficulties with complex semantic structures (e.g., "by whom she had...") and the current lack of support for enumerated lists of children. In a large-scale test, OntoSoar processed an 830-page book in approximately 8 hours, extracting over 35,000 assertions, demonstrating significant throughput for historical research.

Discussion and Future Work

A key insight from this research is that downstream semantic processing can effectively correct upstream syntactic parsing errors by applying domain-specific ontological constraints. Furthermore, the successful application of the Soar architecture to large-scale text processing marks a significant milestone for cognitive modeling.

Future work aims to reduce manual rule engineering by developing a compiler for ECG-to-Soar translation. Additionally, the authors propose "learning by observation," where the agent could adapt to new book styles by observing human annotators, potentially overcoming current limitations in list processing and OCR sensitivity.

Summary updated

BibTeX
@inproceedings{lindes2015ontology,
  author = {Peter Lindes and Deryle W. Lonsdale and David W. Embley},
  title = {{Ontology-Based Information Extraction with a Cognitive Agent}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9227},
}
Figure