Non-Canonical Acts and their Topical Distribution

Christian Vrangbæk, Eva Vrangbæk, Márton Kardos, Kristoffer L. Nielbo, Jacob MortensenSemDH@ESWC 2024 · OpenReview

TL;DR

This paper explores using topic modeling to classify and characterize non-canonical "Acts stories" within New Testament Apocrypha. By applying Non-Negative Matrix Factorization (NMF) to a corpus of ancient Greek texts, the authors demonstrate that this computational approach can reveal nuanced topical distributions, offering a more data-driven alternative to traditional scholarly classifications.

What Problem Does This Paper Address?

The classification of non-canonical texts in New Testament Studies, often referred to as apocrypha, is a complex and debated issue. Current scholarly taxonomies tend to rely on historical classifications shaped by 4th and 5th-century debates about the canon, leading to a binary canonical/non-canonical division. Furthermore, these classifications often use broad conceptual labels like "Gnosticism" or "Encratite," which may oversimplify the rich and varied topical content of these texts. The authors aim to address this by developing a computationally driven classification system for New Testament Apocrypha, using topic modeling to create a structured ontology based on textual data. Specifically, they investigate the topical distribution of four non-canonical "Acts stories" to see if topic modeling can provide a more objective and detailed characterization.

Methods

The core methodology employed is Non-Negative Matrix Factorization (NMF), a topic modeling technique. NMF decomposes a high-dimensional term-document matrix into two lower-rank matrices, representing topics and their distribution within documents, and words and their association with topics.

The process involved several key steps:

  • Text Corpus: A corpus of 2153 Ancient Greek texts was compiled from various sources (Perseus Corpus, First1KGreek, Pseudepigrapha.org, Deutsche Bibelgesellschaft) to provide a relevant historical and literary context for the target texts.
  • Preprocessing: Texts underwent cleaning (removing Latin characters, digits, extra whitespace) and lemmatization using a transformer-based pipeline (OdyCy) built on Ancient Greek BERT and SpaCy.
  • Vectorization: The TF-IDF (Term Frequency-Inverse Document Frequency) method was used to represent textual data as numerical vectors. TF-IDF captures the importance of a term within a document relative to its frequency across the entire corpus.
  • Topic Modeling (NMF): NMF was applied to the preprocessed and vectorized corpus. The authors determined that 10 topics provided the most interpretable coherence for their dataset.
  • Topic Interpretation: Topics were interpreted qualitatively by domain experts, who assigned descriptive labels (e.g., "Historical-political topic," "Element philosophy") based on the top-ranking words within each topic and by examining texts with clear topical distributions.

Datasets

  • Target Texts: Four non-canonical "Acts stories": Acta Joannis, Acta Thomae, Acta Barnabae, and Acta Philippi.
  • Corpus: A collection of 2153 Ancient Greek texts retrieved from the Perseus Corpus, First1KGreek, Pseudepigrapha.org, and Deutsche Bibelgesellschaft. This corpus serves as the contextual background for analyzing the target texts.

Key Insights and Ideas

  • Topic Modeling as a Classification Tool: The study strongly suggests that topic modeling is a viable and valuable tool for classifying and characterizing apocryphal texts, offering a data-driven approach that complements traditional methods.
  • Nuanced Topical Distributions: The four target "Acts stories" exhibit complex and multifaceted topical distributions, distinguishing them from texts with more singular topical dominance. This complexity is precisely what traditional, broader classifications might miss.
  • Bridging Traditional and Computational Methods: The paper emphasizes the importance of a collaborative approach, where computational methods are integrated with traditional scholarly expertise to achieve a deeper understanding of texts.
  • Ontological Framework: The ultimate goal is to use these topical distributions to build a structured ontology for New Testament Apocrypha, facilitating a more systematic and computationally accessible classification system.
  • Visualizing Textual Relationships: The scatterplots generated from the topic modeling provide a visual representation of how texts relate to each other based on their topical content, offering navigational insights into the corpus.

Key Results

  • The four target non-canonical "Acts stories" (Acta Joannis, Acta Thomae, Acta Barnabae, Acta Philippi) showed a generally similar distribution across the dominant Topic 2 (god, people, human) and other theological/philosophical topics.
  • However, the distribution of minor topics revealed significant differences, allowing for a more nuanced characterization of each text. For instance, Acta Joannis and Acta Thomae were placed more securely within the Topic 2 cluster, while Acta Philippi and Acta Barnabae were drawn towards the center, likely due to a higher proportion of Topic 0 (historical-political).
  • The analysis successfully mapped the topical distribution of these texts within the broader corpus, showing how they cluster and relate to other texts based on their thematic content.

Limitations and Future Work

  • Interpretive Subjectivity: While the authors strive for transparency, the qualitative interpretation of topic labels still involves a degree of scholarly judgment.
  • Corpus Dependency: The identified topics and their distributions are inherently dependent on the chosen corpus. The authors acknowledge that they do not have access to all texts that might have influenced the authors of the target texts.
  • Future Work: The authors intend to use these findings to build a more comprehensive New Testament Apocryphal Ontology. They also aim to extend this analysis to include more texts, further refining the classification system and exploring the potential of topic modeling for understanding historical and literary contexts.

Summary updated

BibTeX
@inproceedings{vrangbk2024non,
  author = {Christian Vrangbæk and Eva Vrangbæk and Márton Kardos and Kristoffer L. Nielbo and Jacob Mortensen},
  title = {{Non-Canonical Acts and their Topical Distribution}},
  booktitle = {SemDH@ESWC},
  year = {2024},
  url = {https://openreview.net/forum?id=LYyJcWkz26},
}
Figure