Non-Literal Text Reuse in Historical Texts: An Approach to Identify Reuse Transformations and its Application to Bible Reuse

Maria Moritz, Andreas Wiederhold, Barbara Pavlek, Yuri Bizzoni, Marco BüchlerEMNLP 2016 · DOI

Introduction: The Challenge of Historical Text Reuse

Detecting text reuse in historical Ancient Greek and Latin texts presents a "boss level" challenge for modern Natural Language Processing (NLP). Unlike contemporary plagiarism detection, which benefits from standardized orthography and massive datasets, historical texts often lack common spelling conventions and have evolved stylistically over centuries. Furthermore, they exist in fragmented "lines of transmission"—manuscripts that have diverged significantly over time.

Historical authors rarely "copy-paste." Instead, they weave source material into their own prose, altering grammar (morphology) to fit new contexts or swapping words for synonyms. Standard statistical methods often fail to capture these nuanced adaptations. This paper addresses this gap by developing an automated framework to categorize specific semantic, morphological, and lexical transformations that occur when a source text, such as the Bible, is adapted into a new work.

Methodology: A Multi-Pronged Approach

The researchers proposed a framework combining automated classification with qualitative manual analysis to bridge the gap between literal and non-literal reuse.

Automated Transformation Analysis

The core algorithm compares a reused text excerpt to its original source using a hierarchy of "operations":

  • Literal: Exact matches or simple case changes.
  • Lemmatization: Matching words based on their dictionary headword (e.g., "running" to "run").
  • Semantic: Utilizing the Ancient Greek WordNet (AGWN) to identify replacements by synonyms, hypernyms (broader terms), or hyponyms (narrower terms).
  • Morphological: Identifying changes in Part-of-Speech (PoS) or grammatical case while the root word remains constant.
Examples of literal and non-literal reuse transformations in Ancient Greek and Latin.

Integration of Linguistic Resources

To maximize coverage of ancient vocabulary, the system integrates specialized tools such as the Classical Language Tool Kit (CLTK) and the Biblindex lemma lists.

Manual Expert Review

To establish a performance "ceiling," a Latinist and an Archaeologist manually analyzed a data subset. This helped identify complex transformations that current software misses, such as multi-word expressions or vague semantic allusions.

Datasets

The study utilized two primary datasets where Bible reuse had been previously identified by scholars:

  • Ancient Greek: 199 verse-reuse pairs from Clement of Alexandria’s Salvation for the Rich (2nd Century AD).
  • Latin: 162 verse-reuse pairs from the works of Bernard of Clairvaux (12th Century AD).
  • Source Texts: The Greek Septuagint, the Greek New Testament, and the Latin Vulgate.

Key Insights and Results

The study revealed a significant "literalness gap" between the two languages. Latin reuse (Bernard) was notably more literal than Ancient Greek reuse (Clement). In the Greek dataset, fewer than 25% of tokens overlapped with the original, underscoring why simple keyword searches are insufficient for historical research.

Logarithmic distribution of transformation functions used across different reuse lengths for Greek and Latin.

Automated Performance

The automated approach successfully identified at least one transformation operation for 45% of words in the Greek dataset and 67% in the Latin dataset. The "lemma support" metric ($sup_{lem}$) indicated that resources covered 65% of Greek reuse and 88% of Latin reuse.

Semantic and Morphological Shifts

Manual analysis showed that synonyms accounted for roughly 40% of replacements in both languages. In the Latin data, Part-of-Speech changes represented 15% of all replacements, with frequent transitions from pronouns to nouns and participles to verbs.

Limitations and Future Directions

Currently, the system acts as a "classifier" for known reuse rather than a "detector" for finding new instances. A major bottleneck remains the coverage of existing linguistic databases; many domain-specific or author-specific words are missing from standard WordNets.

Future work aims to use these identified "edit scripts" to train machine learning models for detecting non-literal reuse in massive, un-annotated digital libraries. Furthermore, learning these transformation patterns could assist in "reverse-engineering" lost original texts from their surviving reuses.

Summary updated

BibTeX
@inproceedings{moritz2016non,
  author = {Maria Moritz and Andreas Wiederhold and Barbara Pavlek and Yuri Bizzoni and Marco Büchler},
  title = {{Non-Literal Text Reuse in Historical Texts: An Approach to Identify Reuse Transformations and its Application to Bible Reuse}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2016},
  doi = {10.18653/v1/d16-1190},
}
Figure