LambdaG 0.2.0

The Python library containing the implementation of the authorship verification method 'LambdaG', including a kneser-ney language model implementation.

LambdaG - Grammar as a behavioral biometric: Using cognitively motivated grammar models for authorship verification

This is the official repository for the paper "Grammar as a behavioral biometric: Using cognitively motivated grammar models for authorship verification". The paper proposes an authorship verification (AV) method - called LambdaG - which seeks to answer the question of whether two given documents are written by the same author, or not. In contrast to existing AV methods which often suffer from high complexity, low explainability and especially from a lack of clear scientific justification, LambdaG represents a simpler method based on modeling the grammar of an author following Cognitive Linguistics principles.

Given two documents, 𝒟_𝒜 and 𝒟_𝒰 as well as some reference documents $𝔻_{\text{ref}}$, the ratio of the likelihoods of 𝒟_𝒰 given 𝒜's grammar versus the grammar of a reference population’s grammar is calculated. Based on this ratio a final prediction of whether 𝒜=𝒰 holds is made.

Additionaly, this implementation also exposes two novel normalisation techniques (Square Root Correction and the Hapax Correction), proposed in the paper "Normalisation-Based Likelihood Ratio Estimation for Forensic Authorship Verification", for deriving likelihood ratios from LambdaG without the need of a training dataset.

Usage

Python implementation

To use LambdaG with Python install the LambdaG package from PyPI:

pip install lambdag

or alternatively directly from this repository:

pip install git+https://github.com/AndreaNini/LambdaG

Afterwards you can use LambdaG as follows, for example:

from sklearn.metrics import accuracy_score, roc_auc_score
from lambdag.corpus import load_corpus
from lambdag import LambdaGMethod

# instantiate method object
method = LambdaGMethod(
    basis="tokens", 
    order=10, 
    num_references=30, 
    calibration="logistic_regression"
)

# train method 
# (*not* needed when using `square_root` or `hapax` calibration)
train_problems, train_labels, train_author_texts = load_corpus("corpus/path/train")
method.fit(train_problems, train_author_texts, train_labels)

# evaluate on test corpus
test_problems, test_labels, test_author_texts = load_corpus("corpus/path/test")
test_probas = method.predict_proba(test_problems, test_author_texts)

print(f"Accuracy: {accuracy_score(test_labels, test_probas[:,1]>=0.5):.3f}")

A text heatmap with colour-coded tokens depending on the $\lambda_G$ value can be made using the lambdaG_visualize() function (see example). This feature requires the optional visualization dependencies, which can be installed with pip install lambdag[visualize].

For further examples see examples.

R implementation

An alternative implementation in R is provided in the idiolect package.

Data

This is selection of datasets used in the paper that can be shared freely. Please note that, however, for copyright and GDPR reasons these are released only after pre-processing for content-masking (using the POSnoise algorithm). If you are interested in the other datasets please email us.

Name	Link to repository
Enron corpus
Wiki
Perverted Justice
Apricity
TripAdvisor
Blog

Please keep in mind that, for best results, LambdaG should be used on texts that have been preprocessed with POSNoise (paper, original implementation).

Citation

If you use our code in your work, please consider citing:

@article{nini2025grammarbehavioralbiometricusing,
    title={Grammar as a Behavioral Biometric: {{Using}} Cognitively Motivated Grammar Models for Authorship Verification}, 
    author={Nini, Andrea and Halvani, Oren and Graner, Lukas and Titze, Sophie and Gherardi, Valerio and Ishihara, Shunichi},
    year={2026},
    journal = {Humanities and Social Sciences Communications},
    volume = {13},
    number = {1},
    pages = {455},
    issn = {2662-9992},
    doi = {10.1057/s41599-025-06340-3},
}

and (especially when using square_root and hapax calibration methods):

@misc{barlow2026normalisationbasedlikelihoodratioestimation,
      title={Normalisation-Based Likelihood Ratio Estimation for Forensic Authorship Verification}, 
      author={Sadie Barlow and Andrea Nini and Edoardo Manino},
      year={2026},
      eprint={2607.09501},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.09501}, 
}