Using Bibliodata LODification to Create Metadata-Enriched Literary Corpora in Line with FAIR Principles

Agnieszka Karlinska, Cezary Rosiński, Marek Kubis, Patryk Hubar, Jan WieczorekCOLING 2024

TL;DR

This paper introduces a novel workflow for building metadata-rich literary corpora, focusing on the 19th and 20th-century Polish novel. By leveraging Linked Open Data (LOD) principles and a sophisticated multi-stage metadata enrichment process, they've created the 19/20MetaPNC corpus, designed for enhanced reusability and computational literary studies, addressing limitations of previous approaches.

What Problem Does This Paper Address?

Building comprehensive and useful literary corpora for computational analysis presents several challenges. Unlike linguistic corpora, literary ones need to capture nuances of literary expression, not just linguistic structures. Standard corpus creation methods, often borrowed from linguistics, struggle with the unique demands of literary research, particularly regarding metadata. Existing literary corpora often lack standardized metadata, are not balanced across various dimensions (like genre, period, or author origin), and can be biased towards canonical texts, overlooking less prominent works. Furthermore, the process of collecting, cleaning, and enriching metadata from diverse sources (bibliographic catalogs, digital libraries) is labor-intensive and prone to inconsistencies. This paper aims to bridge the gap between computational linguistics and digital humanities by proposing a robust, FAIR-aligned methodology for creating literary corpora that are both research-ready and highly reusable.

Methods

The core of the paper's methodological contribution lies in its approach to metadata enrichment and corpus design, guided by the FAIR (Findable, Accessible, Interoperable, Reusable) principles and the 7-star Linked Data Service standards.

Key methods include:

  • Bibliodata LODification: This involves transforming bibliographic data from traditional catalogs into Linked Open Data (LOD) using Resource Description Framework (RDF). This allows for machine-readable and interoperable metadata.
  • Multi-stage Metadata Enrichment: Beyond standard bibliographic data, the process incorporates:
  • NLP-based Enrichment: Named Entity Recognition (NER) is used to identify and tag geographical locations within the novels.
  • Toponym Disambiguation: A four-stage workflow is employed to accurately identify and standardize geographical entities, linking them to historical context and partitions.
  • PID and LOD Enrichment: Persistent Identifiers (PIDs) are assigned to entities (authors, places) to ensure unique identification and facilitate cross-referencing with external LOD repositories like VIAF and Wikidata.
  • Research Question-Driven Design: The corpus design is explicitly driven by specific research questions (e.g., the urban-rural dichotomy in Polish fiction), ensuring the metadata collected is relevant to these inquiries.
  • Sophisticated Balancing Criteria: The corpus is balanced not only by literary period and genre but also by author gender, geographical origin (partitions), and reception level (number of reprints).
  • Knowledge Graph Representation: The final metadata is stored and presented as a knowledge graph (using Neo4J and RDF), enabling complex querying and visualization.

Datasets

The paper describes the creation of the 19/20MetaPNC (Metadata-enriched Polish Novel Corpus from the 19th and 20th centuries). The initial dataset was compiled from four open sources:

  • ELTeC corpus (ELTEC, 2021): Novels encoded in TEI format.
  • Wolne Lektury (Modern Poland Foundation, 2022): An online repository of school readings.
  • Wikisource (Wikimedia Foundation, 2022): Transcriptions of public domain books.
  • Polona (National Library of Poland, 2022): Scans of printed books with OCR layers.

The initial collection comprised approximately 1,707 unique novels after deduplication, from which 1,000 were randomly selected based on the defined balancing criteria. No new benchmarks were introduced, but the paper discusses how the 19/20MetaPNC corpus relates to and improves upon existing literary corpora like ELTeC and KorBa.

Key Insights and Ideas

  • Metadata is Key for Reusability: The paper strongly emphasizes that well-structured, enriched metadata is crucial for making literary corpora useful beyond their initial research purpose, aligning with FAIR principles.
  • LOD is a Game-Changer for Literary Corpora: Transforming bibliographic data into LOD is presented not just as a technical step but as a fundamental enabler for interoperability, discoverability, and advanced analysis of literary texts.
  • Balancing is More Than Just Genre/Period: The authors demonstrate that true balance in literary corpora requires considering multiple dimensions, including geographical origin (partitions), author demographics, and reception history, to avoid biases.
  • Bridging the Humanities-Linguistics Divide: The project successfully integrates techniques from computational linguistics (NLP) with the specific needs and data sources of literary studies and digital humanities.
  • Knowledge Graphs Unlock Deeper Insights: Representing the corpus as a knowledge graph allows for complex relational queries and visualizations that reveal patterns and connections not easily discoverable in traditional flat file formats.

Key Results

  • Creation of the 19/20MetaPNC Corpus: A balanced corpus of 1,000 Polish novels from 1854-1939, featuring multi-dimensional balancing criteria.
  • Comprehensive Workflow: A nine-stage workflow for literary corpus creation and reuse is proposed, integrating research design, data collection, NLP enrichment, LODification, and publishing.
  • Metadata Enrichment: The corpus metadata includes standard bibliographic data, NLP-derived geographical information, author PIDs, and place PIDs, all linked to external LOD sources.
  • Quantitative Corpus Statistics: The corpus contains over 64 million tokens and 4.2 million sentences. The paper provides detailed statistics on token and sentence counts across various balancing criteria (gender, partition, period, reception). For example, women authors constitute 29.4% of the corpus, and novels with low reception (≤2 reprints) make up 63.3%.
  • Demonstrated Utility: A Neo4J graph query visualization is presented, showing the relationship between book publication location and author's birth partition, highlighting disparities and transfer of literary texts between partitions.

Limitations and Future Work

  • Opportunistic Corpus: The authors acknowledge that the corpus is "opportunistic" due to the reliance on existing digital texts, meaning it may not perfectly represent the entire historical population of Polish novels.
  • Data Quality Challenges: The reliance on OCR-derived texts and the inherent limitations of historical bibliographic data presented significant challenges in data cleaning and metadata validation.
  • Future Work:
    • Expansion: Include historical novels with plots predating 1815 and rebalance the corpus based on novel length.
  • Further NLP Enrichment: Determine thematic content using supervised/unsupervised approaches and crowdsourced datasets.
  • Text Corpora Ontology (TCO) Development: Create a comprehensive ontology for publishing text corpora in a Semantic Web environment, integrating existing ontologies and representing corpus attributes like balance and representativeness.
  • Broader Application: Generalize the proposed workflow for other types of literary research and corpora.

Summary updated

BibTeX
@inproceedings{karlinska2024using,
  author = {Agnieszka Karlinska and Cezary Rosiński and Marek Kubis and Patryk Hubar and Jan Wieczorek},
  title = {{Using Bibliodata LODification to Create Metadata-Enriched Literary Corpora in Line with FAIR Principles}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure