Ontology-based design information extraction and retrieval

Zhanjun Li, Karthik RamaniArtif. Intell. Eng. Des. Anal. Manuf 2007 · DOI

TL;DR

This paper introduces an ontology-based approach to automatically extract and retrieve design information from unstructured textual documents. By leveraging domain-specific ontologies and shallow natural language processing, the system creates structured representations of design semantics, significantly outperforming traditional keyword-based search for design retrieval.

What Problem Does This Paper Address?

The increasing complexity of modern products and the proliferation of digital design documentation have created a massive challenge for engineers: how to effectively find and reuse relevant design information. Existing Product Data Management (PDM) and Product Lifecycle Management (PLM) systems often rely on manual metadata indexing, which is time-consuming and prone to errors. Engineers spend a significant portion of their time searching for information, hindering design exploration, learning, and reuse. This paper addresses the need for a more intelligent and automated system that can understand the semantic content of design documents, not just keywords, to facilitate efficient retrieval.

Methods

The core of the proposed system, named ODART (Ontology-based Design document Analysis and Retrieval Tool), involves two main stages: design semantics extraction and ontology-based query processing.

Design Semantics Extraction: 1. Shallow NLP: The system employs techniques like tokenizing, Part-of-Speech (POS) tagging, and phrase chunking to analyze input English text. This is a less computationally intensive approach than deep NLP, focusing on extracting relevant design concepts. 2. Domain Ontology: A crucial component is a layered domain ontology. The "domain layer" contains generalized concepts and taxonomies (e.g., device, function, material) applicable across a domain, while the "application layer" is company or product-specific and is automatically generated from design documents. This ontology guides the extraction process by providing a structured vocabulary and relationships. 3. Concept Recognition & Relationship Extraction: Phrases identified by NLP are matched against concepts in the domain ontology. A scoring mechanism (Tscore and Cscore) is used to select the most appropriate concept. Relationships between concepts (e.g., "has_function," "has_part") are identified using semantic rules and the ontology's structure. 4. Application Ontology Generation: The extracted concepts and relationships from a document are integrated to form a concept graph, which then populates the application-specific ontology for that document. This creates a structured, semantic representation.

Ontology-based Query Processing: 1. Syntactic Metric: User queries are analyzed using the domain ontology to identify relevant concepts. Concept propagation (selecting child concepts of a matched parent) is used to broaden recall. 2. Semantic Metric: To improve precision, the system uses concept pairs (related concepts within the application ontology) to disambiguate queries and prune irrelevant results. This metric measures semantic closeness between the query and document representations.

Datasets

The study utilized design reports from a senior engineering design class (ME444) at Purdue University. This dataset comprised 158 design reports collected over three years, each containing summary sections describing the physical prototype and manufacturing. The average summary section length was 16.26 sentences with an average of 12.37 words per sentence.

Key Insights and Ideas

  • Semantics over Keywords: The fundamental insight is that understanding the meaning and relationships within design documents, rather than just matching keywords, is critical for effective retrieval.
  • Ontology as a Bridge: Domain ontologies act as a crucial bridge, providing a structured vocabulary and semantic framework to interpret the nuances of natural language in engineering contexts.
  • Layered Ontology Approach: The distinction between a general "domain ontology" and a specific "application ontology" allows for both broad applicability and precise representation of individual designs.
  • Concept Pairs for Disambiguation: Using concept pairs (linked concepts in the application ontology) is a novel way to resolve semantic ambiguity in queries, leading to more accurate retrieval.
  • Automated Extraction: The system's ability to automatically extract and structure design semantics reduces the manual effort typically required for knowledge base construction.

Key Results

The paper reports significant improvements in retrieval performance compared to the traditional vector model (keyword-based search).

  • Average Recall: The ontology-based model achieved an average recall of 93%, compared to 60% for the vector model.
  • Average Precision: The ontology-based model achieved an average precision of 89%, compared to 82% for the vector model.
  • General Queries: The ontology-based approach showed particularly strong performance for general queries (e.g., "search for a vehicle"), achieving 100% recall and precision, due to concept propagation.
  • Specific Queries: While the gains were less dramatic for highly specific queries, the ontology-based method still outperformed the vector model in most categories.

Limitations and Future Work

  • Ontology Maintenance Cost: The authors acknowledge that maintaining and extending domain ontologies with new concepts can be costly. They suggest exploring ontology learning techniques from large corpora.
  • Scope of Documents: The current prototype primarily uses design project reports. Future work aims to expand processing to other document types like project proposals, memos, and emails to cover the entire product lifecycle.
  • Informal Documents: Extracting semantics from less formal documents like engineers' logbooks, which may not adhere to standard documentation formats, presents a challenge.
  • Shape-Based Retrieval Integration: The authors plan to integrate their textual retrieval system with shape-based engineering part retrieval to achieve more comprehensive search results.
  • Larger Datasets and Environments: Future work includes evaluating the system's performance on larger datasets and in different design environments.

Summary updated

BibTeX
@article{li2007ontology,
  author = {Zhanjun Li and Karthik Ramani},
  title = {{Ontology-based design information extraction and retrieval}},
  journal = {Artif. Intell. Eng. Des. Anal. Manuf.},
  year = {2007},
  doi = {10.1017/s0890060407070199},
}
Figure