This repository contains the input data, scripts and Suppl. Tables and Figures from a study that employs distributional linguistics, a newly introduced technique to examine the controversial concept of verbal ‘deponency’ in LXX-NT (koine) Greek.
Renaissance grammarians introduced ‘deponency’ as an additional classifier to the traditional voice paradigm (active, middle, passive) to account for verbs exhibiting middle-passive morphology with active interpretation.
While contemporary New Testament scholars often dismiss this classification as superfluous, advocating for its reintegration into the middle voice category, our distributional analysis of the LXX-NT corpus reveals distinct grammatical patterns.
The analyses demonstrates that ‘deponent’ verbs inhabit grammatical contexts more closely aligned with active voice rather than middle voice forms.
These findings suggest that Renaissance grammatical insights merit reconsideration, potentially warranting a reassessment of deponency’s role in Greek verbal taxonomy rather than its wholesale rejection from contemporary linguistic analysis.
The study employs distributional analysis, a quantitative linguistic method, to investigate the debated concept of deponency in LXX-NT Greek. Traditional grammatical approaches have long struggled with the classification of deponent verbs—those that exhibit middle or passive morphology but are interpreted with active meaning. While contemporary scholars often advocate for the elimination of deponency as an unnecessary Latin-based construct, this study challenges that position by utilizing corpus-wide computational analysis to detect systematic grammatical patterns.
Historically, the concept of deponency was introduced by Renaissance grammarians who sought to impose Latin grammatical structures onto Greek, leading to the classification of certain middle-voice verbs as 'deponent'. Modern scholarship, particularly from the late 20th and early 21st centuries, has called this classification into question, arguing that the Greek middle voice itself is poorly understood and that many so-called deponent verbs may be better analyzed as authentic middle forms. However, this critique has largely been based on qualitative and interpretive analyses, without large-scale empirical validation.
The present study fills this gap by applying distributional analysis, a method rooted in computational linguistics, to the entire LXX-NT corpus. Unlike traditional approaches that emphasize lexical meaning, this study systematically isolates grammatical form by constructing an artificial corpus where lemmas are replaced with their corresponding grammatical categories. This approach enables an objective examination of the syntactic behavior of deponent verbs without interference from semantic interpretations.
The core of the methodology involves generating vector-based word embeddings using a modified GLoVe model, a well-established distributional approach that quantifies word similarity based on co-occurrence patterns. The study specifically measures whether deponent verbs cluster more closely with active-voice or middle-voice grammatical forms. The findings indicate that in 75.4% of cases, deponent verbs exhibit stronger syntactic similarity to active voice constructions than to middle voice constructions. This statistically significant pattern suggests that Renaissance grammarians may have correctly identified a real grammatical phenomenon rather than erroneously importing Latin concepts onto Greek.
The GLoVe (Global Vectors for Word Representation) model is a widely used unsupervised learning algorithm for generating word embeddings based on statistical co-occurrence information across a large corpus. Unlike models such as Word2Vec, which rely on predictive learning frameworks, GLoVe constructs embeddings based on the global co-occurrence matrix, ensuring that words appearing in similar contexts are mapped to similar vector representations.
GLoVe operates by first constructing a word co-occurrence matrix, where each entry represents the frequency with which two words appear together within a defined context window. This co-occurrence data is then transformed into a log-bilinear model, which optimizes vector representations by minimizing the difference between predicted and actual co-occurrence probabilities. The key advantage of this approach is that it preserves both local and global corpus information, offering a more comprehensive representation of linguistic structures.
Mathematically, the GLoVe model solves the following optimization problem:
where:
-
$X_{ij}$ represents the co-occurrence count between words$i$ and$j$ -
$w_i$ and$w_j$ are the word vectors being trained -
$b_i$ and$b_j$ are bias terms -
$f(X_{ij})$ is a weighting function to balance frequent and infrequent co-occurrences
By factoring the logarithm of co-occurrence probabilities, GLoVe captures meaningful semantic relationships between words. The model assigns similar vector representations to words that frequently co-occur in similar contexts, thereby encoding syntactic and semantic similarities within the embedding space.
In the context of this study, the LXX-NT corpus was processed using the GLoVe framework, ensuring that grammatical forms were embedded in a continuous vector space based purely on their distributional behavior. The selected context window of five words on either side of the target word allowed the model to capture morpho-syntactic dependencies, effectively distinguishing active, middle, and deponent forms based on their statistical relationships.
By leveraging GLoVe embeddings, this study provides an empirical, corpus-wide validation of deponency as a potential grammatical category, countering the prevailing scholarly consensus that it should be discarded. The approach also highlights the importance of quantitative methods in addressing long-standing linguistic debates. By prioritizing grammatical distribution over semantic intuition, this study provides a reproducible, data-driven framework for analyzing voice distinctions in LXX-NT Greek. Future research could refine these findings by integrating probabilistic modeling to express deponency as a graded rather than binary category, potentially reconciling traditional and modern perspectives on Greek voice morphology.
For an in-depth understanding of vector semantics and embeddings, check Chapter 6 of Speech and Language Processing by Jurafsky and Martin 1. Additionally, Wishart's work provides a readable application of vector embeddings in Koine Greek semantics 2.
Footnotes
-
Jurafsky, D., & Martin, J. H. (2025). Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition with Language Models (3rd ed.). Retrieved from https://web.stanford.edu/~jurafsky/slp3/6.pdf ↩
-
Wishart, R. A. (2018). Hierarchical and Distributional Lexical Field Theory: A Critical and Empirical Development of Louw and Nida’s Semantic Domain Model. International Journal of Lexicography, 31(4), 394–419. https://doi.org/10.1093/ijl/ecy015 ↩