FabKG: A Knowledge graph of Manufacturing Science domain utilizing structured and unconventional unstructured knowledge source

Aman Kumar, Akshay Bharadwaj, Binil Starly, Collin LynchNAACL 2022

TL;DR

This paper introduces FabKG, a novel knowledge graph for the manufacturing domain that integrates structured data from Wikidata with unstructured knowledge from textbooks and student notes. This approach aims to create a more comprehensive and accessible knowledge base for manufacturing science, enabling advanced applications like domain-specific question answering and formula-based calculations.

What Problem Does This Paper Address?

The rapid growth of information in manufacturing science has led to a vast amount of knowledge, but much of it remains unstructured and difficult to access. While knowledge graphs (KGs) are powerful tools for organizing and utilizing information, existing KGs often lack the depth and specificity required for the manufacturing domain. This gap hinders intelligent manufacturing processes, adaptive education, and efficient decision-making. The authors highlight that even comprehensive resources like Wikidata have limitations in representing the nuanced terminology and specific entities within manufacturing. Furthermore, valuable knowledge embedded in student notes and textbooks is often underutilized.

Methods

The authors propose a multi-pronged approach to construct FabKG:

  • Leveraging Structured Data (Wikidata): They extract relevant entities and relations from Wikidata using two methods:
  • Vocabulary-based: Utilizing textbook index words, research paper keywords, and named entity recognition (NER) results from their FabNER system to query Wikidata.
  • Unstructured Text-based: Employing a semi-supervised approach that uses student notes, with standard textbooks as a reference.
  • Extracting from Unstructured Sources:
    • Textbook Index Words: Extracting entities from the index sections of five diverse manufacturing ebooks.
  • Research Paper Keywords: Collecting keywords from over 500,000 manufacturing abstracts.
  • Named Entity Recognition (NER): Using a BERT-based BiLSTM-CRF model to identify entities in review articles related to manufacturing.
  • Integrating Student Notes: A novel crowdsourcing method where students' exam cheatsheets/notes are used. These notes, often containing crucial information for learning, are structured using specific formatting guidelines to facilitate extraction of entities and relations.
  • Fusion and Merging: Combining the extracted structured and unstructured knowledge into a unified KG. They address the challenge of entity linking by using synonyms and identifying links that might not be explicitly present in Wikidata.
  • Knowledge Graph Construction: The process involves fetching Wikidata items, extracting subgraphs, and then merging them with knowledge derived from textbooks and student notes. They also use SPARQLWrapper to traverse the Wikidata graph and fetch related entities.

Datasets

  • Wikidata: Used as a primary source of structured knowledge.
  • Manufacturing Textbooks: Five ebooks were used to extract index entities.
  • Research Paper Abstracts: Over 500,000 abstracts were processed for keywords.
  • Review Articles: Ten full review articles related to manufacturing were used for NER.
  • Student Notes/Cheatsheets: A novel source of unstructured knowledge, collected from students.

The paper introduces a merged knowledge graph, FabKG, which combines these sources.

Key Insights and Ideas

  • Bridging the Structured-Unstructured Divide: The core idea is to effectively combine the richness of structured knowledge bases like Wikidata with the often-overlooked, domain-specific details found in unstructured sources like textbooks and student notes.
  • The Value of Student Notes: Student notes, typically discarded after exams, represent a valuable, human-curated source of essential knowledge that can be leveraged for KG construction. This is a form of indirect crowdsourcing.
  • Comprehensive Manufacturing Knowledge: Existing KGs are often too general. FabKG aims to capture fundamental manufacturing concepts, processes, parameters, materials, and applications, making it a more specialized and useful resource.
  • Enabling Advanced QA: By incorporating formulas and expressions as entities and relations, FabKG can go beyond simple fact retrieval to perform inference-based calculations, mimicking human reasoning.

Key Results

  • FabKG Size: The constructed knowledge graph contains over 65,000 triples.
  • Entity Extraction:
    • ~3500 relevant entities from textbook indexes.
  • ~4500 relevant entities from research paper keywords.
  • ~2000 entities from review articles using NER.
  • Over 1200 distinct entities and 4200 unique triples extracted from student notes.
  • NER Performance: The BERT-based NER model achieved an F-score of 83% for 12 entity categories.
  • Use Cases Demonstrated: The paper showcases FabKG's utility for:
  • Domain-specific question answering (e.g., "Which tool geometry is used for planning?").
  • Expression/formula-based question answering (e.g., calculating strain given force, area, and Young's modulus).

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that Wikidata has a limited amount of specialist terminology for the manufacturing domain, necessitating the integration of other sources.
  • The semi-supervised approach for student notes relies on specific formatting, which might not be universally adopted.
  • Determining directed relations between multiple entities in a sentence can be challenging.
  • Future Work:
    • Directed Relation Prediction: Applying this to larger corpora and performing graph embedding.
  • Utilizing Lecture Presentations: Exploring lecture slides for entity and relation extraction.
  • Entity-Attribute-Value Graphs: Expanding the KG to include attribute-value pairs (e.g., material strength, carbon content) and potentially using hypergraphs for multimodal data.
  • Global Knowledge Integration: Linking FabKG to global knowledge bases like Wikidata using tools like Wikimapper.

Summary updated

BibTeX
@inproceedings{kumar2022fabkg,
  author = {Aman Kumar and Akshay Bharadwaj and Binil Starly and Collin Lynch},
  title = {{FabKG: A Knowledge graph of Manufacturing Science domain utilizing structured and unconventional unstructured knowledge source}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure