index.qmd

---
title: "Authorship Attribution with Machine Learning"
editor: visual
author: Sam McDowell
date: "12-05-2025"
date-format: long
format:
  html:
    toc: true # includes table of contents
    code-fold: true # option for collapsing code blocks in html
execute:
  echo: true # includes output
  warning: false # turns off warnings
  error: false # if set to true, then stops running at error
  output: true
---

Honors Petition Project

Liberty University - School of Business

[Github Link](https://github.com/sammcdo/ML-Authorship-Attribution)

## Introduction

"For as long as universities have existed, one rule has stood firm: a student's work should reflect their own understanding...What happens when the 'author' of the borrowed words isn't a person at all, but an algorithm?" Nneoma Agwu-Okoro raised this important question in a recent article on AI and plagiarism (2025). She goes on, "not long ago, a philosophy professor in the US was left impressed by what seemed to be a brilliantly argued essay... He later discovered it had been written almost entirely by AI." (2025) Identifying the author of a written work is more important than ever before.

This creates opportunities to bring the skill sets of modern science and engineering to bear on the problem. Identifying authorship is not a new idea. There are multiple historical examples of authors using pseudonyms to publish their works. In looking for ways to find the real penmen responsible, the study of stylometry was born. In 1964, Professors David Wallace and Frederick Mosteller first used tools from their background in statistics to determine the authorship of written works (Mercer, 2017). They called their new collection of techniques stylometry (Mercer, 2017).

Stylometry is the use of statistical methods to analyze linguistic style (Wermer-Colman, 2023). A unique linguistic style can be used to determine authorship. As questions about the authenticity of written works are on the rise, this is practical opportunity to apply these ideas and explore the best methods of determining authorship. Additionally, as techniques in Machine Learning become more available and more accessible, it opens the opportunity to use the textual features extracted by stylometric methods for use in Machine Learning models.

This petition demonstrates several methods for determining authorship using popular Machine Learning tools. It uses various types of classification to determine authorship from a pool of works.

## Dataset

The dataset used for this demonstration is the Blog Authorship Corpus (Schler et al., 2006). The corpus itself has over six hundred thousand posts, all over 200 words long. This resource is well established in the authorship attribution field, being used as a test source in foundational works such as the creation of the BertAA model (Fabien et al., 2020).

This dataset was then filtered to narrow down the number of authors being used in this experiment. The 15 authors with the highest total number of posts over 1000 words were selected. Each of these authors then had their top 40 longest works extracted. The works were then shortened to 1000 words each. This left the final dataset with 600 blog posts written by 15 distinct authors and each work being 1000 words long.

The texts were also cleaned of their HTML embeds, 'words' made up only of symbols, and the word urlLink from cleaning of the data by the original authors.

```{python}
import pandas as pd
filename = "./data/preprocessed.csv"
df = pd.read_csv(filename, index_col=0)
df.head(n=1)
```

## Feature Engineering

The models used in this work rely on stylometric features extracted from the raw texts. These are split into several categories: vocabulary, readability, and tagging.

### Vocabulary

The vocabulary features start with the Type Token Ratio (TTR). This statistic was first used in the 1950s and has since become a staple in textual features (Cunningham and Haley, 2020). This is the ratio of the number of unique words divided by the total words in the text. It is a well-accepted measure of vocabulary diversity (Cunningham and Haley, 2020). Additionally, the raw number of unique words is included in the features.

The next feature is the stop word ratio, or the ratio of stop words to total words. This was introduced in the foundational work in identifying the authors of the anonymous Federalist Papers in 1966 (Kendall et al.) and continues to be in use today. It is still popular because it is "immune to topic bias" (Koppel et al., 2011).

The next feature is the Hapax Legomena: the number of words used only once in the text (Mardaga, 2012). The name is such because this metric has existed since before the time of Alexander the Great (Mardaga, 2012). The number of words used only twice, or Dislegomena, are also recorded (Mardaga, 2012).

### Readability

The first measure of readability is the average number of syllables per word. This is an important part of many modern readability measures. It was suggested as a primary measure of readability for the military in 1975 (Kincaid et al.).

Sentence length is another very common measure of an author's style (Holmes, 1998). The average and standard deviation of the length are included in the dataset. Average word length and word length standard deviation are also commonly used and included in the dataset (Stamatatos, 2009).

Additionally, there are several standard readability formulas that are included in the dataset. The first is the Flesch Readability Score, which measures the difficulty and complexity of reading a set of text (Eleyan et al., 2020). While this formula has been in use since the early days of stylometry, it is still a standard in readability measurement (Jindal and MacDermid, 2017). It is based on average sentence length and number of syllables (Eleyan et al., 2020).

The second readability calculation is the Dale-Chall Readability Score. This score is supposed to output the suggested grade level of a written work (Gencer, 2024). This formula uses a base list of words deemed suitable for 4th or 5th grade readers and compares how many words are harder than what is on the list (Gencer, 2024). The corpus is taken from ReadabilityFormulas.com (2025).

The final readability measure included in the dataset is the Gunning Fog Index. This measurement marks words as difficult based on their syllables and the number of words per sentence (Plotnikov, 2020). Although developed in the 1940s (Gunning, 1969), it is popular in small sample authorship analysis (Goh et al., 2007).

### Tagging

The final section of data in the dataset is counts of tagged items in the text. The first set of tag counts included are parts of speech. These are very important because each author has their own unique use of sentence structure and phraseology that can be seen here (Gholami-Dasgerdi and Feizi-Derakhshi, 2021). For example, part of speech tagging is all that is needed to classify verse and prose (Chen et al., 2024).

The next group of tagged items are counts of tagged punctuation types. These include periods, exclamations and questions. This is a very common and useful feature that is used for differentiating writing style (Stamatatos, 2009).

The final category of tagged items is sentence structure. Each sentence is categorized as simple, compound, complex or compound-complex. This has been used very effectively in previous work (Feng et al., 2012).

### Overview

The features that resulted can be seen in the following correlation heat map. The features do not strongly correlate to the author of the blog so there is no data leaking through in any of the features. This also means that a more complex model will have the ability to link the non-linear relationships and feature interactions taking place under the surface.

![](output/correl.png)

## Results

### BERT

The first model used for Authorship Attribution is BERT. The BERT Model is a very popular transformer model that was built to understand text deeply (Fabien et al., 2020). A common method for using BERT for classification is to append an additional layer for classification and train it for its new task (Fabien et al., 2020). This implementation is a dense layer on top of the final pooling of the transformer (Hugging Face, 2019).

This model achieved 87% accuracy. The confusion matrix for this model can be seen below.

![](output/BERT_classification.png){width="560"}

### Centroid-Based Authorship Attribution

The second model based on pre-trained models is Centroid-Based Authorship Attribution. In this method, embeddings from a pre-trained model (MPNET Base v2 in this case) are used as a means of classification (Sohrab et al., 2015). To do this, the training texts are embedded and then a mean of the training texts is found to create a centroid (Sohrab et al., 2015). The test texts are then embedded and selected based on which centroid it is closest to (Sohrab et al., 2015). This is done with cosine similarity (Rossiello et al., 2017).

This model achieved 76% accuracy. The confusion matrix for this model can be seen below.

![](output/CosSim_classification.png){width="560"}

### KNN Clusters

K-Nearest Neighbors (KNN) is a standard classification Machine Learning algorithm for labeled data. In this case, the feature set engineered from the texts was used to train a KNN model. KNNs have been used successfully in authorship attribution in previous studies (Luyckx and Daelemans, 2008).

This model achieved 75% accuracy. The confusion matrix for this model can be seen below.

![](output/KNN_classification.png){width="560"}

### Random Forest Classifiers

Given a text feature set, it can be very effective to use a Random Forest Classifier to select the author (Khonji, et al., 2015). In this case, it was very easy to get high accuracy. This was a very effective model in terms of complexity for accuracy trade off.

This model achieved 92% accuracy. The confusion matrix for this model can be seen below.

![](output/RF_classification.png){width="560"}

### Neural Network Classification

Neural Networks are also a common method for achieving accurate classification of authors based on stylometric feature sets (Jafariakinabad and Hua, 2020). In this implementation, 3 layers of Dense nodes with Batch Normalization and Dropout are followed by a final classification Dense layer. This model definitely seems to be suffering from a lack of training data. The batch size is very low because of the low number of samples in the training set.

This model achieved 92% accuracy. The confusion matrix for this model can be seen below.

![](output/NN_classification.png){width="560"}

## Discussion

The results achieved by these models were impressive. There were two different model types that reached 90% accuracy in authorship attribution. However, each model likely has room for improvement.

### BERT

The BERT model most likely needed more training data. This would have included more samples and fewer classes. From how quickly it generalized, it seemed to be learning very well. It seemed to need more data to generalize. However, it was an effective baseline to evaluate the remaining models.

The BERT model is a very popular choice as a base for any textual model because it automatically incorporates syntactic, semantic, and world knowledge (Rogers et al., 2020). BERT representations are known to follow a primitive form of syntactic structure automatically (Rogers et al., 2020). This means the embeddings can include information about parts of speech, phrasing and word roles (Rogers et al., 2020). Some research has shown that while it picks up on this information in the data, it seems not use it to make decisions, so it is likely incomplete (Rogers et al., 2020). Additionally, BERT has semantic knowledge, and it has learned frequent relationships between words in english usage (Rogers et al., 2020). BERT also has some knowledge of interactions commonly described between real world objects (Rogers et al., 2020). In some studies it can compete with knowledge base type implementations (Rogers et al., 2020). This variety of information it has learned is why it is often used as a starting point for so many linguistic analysis tasks.

### Embedding Centroid

The Embedding Centroid method was not effective enough for continued use. It may have use in comparison of two authors but it seemed to get confused by the large number of classes.

Word embeddings are typically trained from a large set of unlabeled data (Kenter et al., 2016). Thus embeddings tend to be general purpose and usable for many tasks (Kenter et al., 2016). The idea of averaging the word embedding to get a sentence embedding is surprisingly effective and supported by literature (Kenter et al., 2016). Additionally, pooling the embeddings into a fixed size representation of a text has been used effectively (Sultana et al., 2024). The implementation used in this example is simpler than these implementations and suffered from inability to generalize because taking the mean embedding lost information or confused information needed for classification.

### K-Nearest Neighbors

The KNN classification method consistently under performed, despite training. Different distance metrics and neighbors were ineffective. Dimensionality reduction with Principal Component Analysis was also ineffective. It might have been suffered from having too many irrelevant features that it could not filter out on its own.

Another problem with the KNN approach can be seen in the visualization of the results with a Principal Component Analysis. This shows how the authors can have very different styles between documents. While the other models are able to learn what features are informative, the difference in style of some texts seems to make nearest neighbor style methods ineffective. This likely applied to the Embedding Centroid method to a lesser extent as well.

### Random Forest

The Random Forest classification worked excellently. It is much less affected by irrelevant data points. This might be helping it extract the most information from these models. A future improvement could be using boosted trees to see if the accuracy could be improved more.

### Neural Network

The final model tested was a Neural Network of the features in the dataset. It was the most successful at classifying the authors. This is important because it is the most flexible for future applications. Compared to the baseline BERT model, this model was 7% more accurate in classification. This could be easily adapted for more authors and more data to create a very robust authorship attribution method.

The Neural Network approach could likely be improved by including a fixed sized pool from the embeddings (Sultana et al., 2024) of the text to incorporate some information about the actual words used in contrast to statistics about the words. This would allow some representation of the content of the document to be included in the Neural Network without needing a large vocabulary and the space complexity of including word counts in the dataset. BERT embeddings could be used for this, as well as other models and embedding types (Sultana et al., 2024).

### Future Research

Overall, the use of stylometric features was comparable to models that used raw text as inputs. Because they are summary statistics, the stylometric features extracted from the text necessarily do not have as much data as is contained in the raw text. However, they also outperformed the methods based on the raw text. To some extent this seems to be because they focused the models on the information that was most unique in the text. However, the results also implies that the accuracy of the feature set based models could be improved by using a form of generalized text information that would allow the models to learn a larger variety of syntactic and semantic quirks of the authors that cannot be seen in general stylometric analysis.

## Conclusion

This petition has shown that authorship classification is a simple task for Machine Learning. The results strongly showed that authorship can be determined both from raw textual input to a model and by features extracted from the texts. The feature engineering done for this dataset could be applied to new texts for a similar set up. The small dataset consisting of only 40 data points per class is a reasonable simulation of a real-world small dataset use case.

As AI authored works are on the rise, it is hopeful that Machine Learning can help identify writing patterns that make our writing creative, unique and individual to each of us.

## References

Adem Gencer. (2024). Readability analysis of ChatGPT’s responses on lung cancer. *Scientific Reports*, *14*(1). https://doi.org/10.1038/s41598-024-67293-2

Agwu-Okoro, N. G. (2025, September 12). *The AI DILEMMA IN HIGHER EDUCATION: Balancing Innovation and Authorship*. Substack.com; Legal Bytes. https://legalbytes.substack.com/p/the-ai-dilemma-in-higher-education

Cunningham, K. T., & Haley, K. L. (2020). Measuring Lexical Diversity for Discourse Analysis in Aphasia: Moving-Average Type–Token Ratio and Word Information Measure. *Journal of Speech, Language, and Hearing Research*, *63*(3), 710–721. https://doi.org/10.1044/2019_jslhr-19-00226

Eleyan, D., Othman, A., & Eleyan, A. (2020). Enhancing Software Comments Readability Using Flesch Reading Ease Score. *Information*, *11*(9), 430. https://doi.org/10.3390/info11090430

Fabien, M., Villatoro-Tello, E., Motlícek, P., & Parida, S. (2020). BertAA : BERT fine-tuning for Authorship Attribution. *International Conference on Networks*, 127–137.

Feng, S., Banerjee, R., & Choi, Y. (2012). Characterizing Stylistic Elements in Syntactic Structure. *ACL Anthology*, 1522–1533. https://aclanthology.org/D12-1139/

Gholami-Dastgerdi, P., & Feizi-Derakhshi, M.-R. (2021). Part of Speech Tagging Using Part of Speech Sequence Graph. *Annals of Data Science*. https://doi.org/10.1007/s40745-021-00359-4

Goh, O. S., Fung, C. C., Depickere, A., & Wong, K. W. (2007). Using Gunnnig-Fog Index to Assess Instant Messages Readability from ECAs. *CiteSeer X (the Pennsylvania State University)*. https://doi.org/10.1109/icnc.2007.800

Gunning, R. (1969). The Fog Index After Twenty Years. *Journal of Business Communication*, *6*(2), 3–13. https://doi.org/10.1177/002194366900600202

Holmes, D. I. (1998). The Evolution of Stylometry in Humanities Scholarship. *Literary and Linguistic Computing*, *13*(3), 111–117. https://doi.org/10.1093/llc/13.3.111

Hugging Face. (2019). *BERT*. Huggingface.co. https://huggingface.co/docs/transformers/v4.37.1/en/model_doc/bert#transformers.BertForSequenceClassification

Jafariakinabad, F., & Hua, K. A. (2020). *A Self-supervised Representation Learning of Sentence Structure for Authorship Attribution*. ArXiv.org. https://arxiv.org/abs/2010.06786

Jindal, P., & MacDermid, J. (2017). Assessing reading levels of health information: uses and limitations of flesch formula. *Education for Health*, *30*(1), 84. https://doi.org/10.4103/1357-6283.210517

Kendall, M. G., Mosteller, F., & Wallace, D. L. (1966). Inference and Disputed Authorship: The Federalist. *Biometrics*, *22*(1), 200. https://doi.org/10.2307/2528232

Kenter, T., Borisov, A., & de Rijke, M. (2016). *Siamese CBOW: Optimizing Word Embeddings for Sentence Representations*. ArXiv.org. https://arxiv.org/abs/1606.04640

Khonji, M., Iraqi, Y., & Jones, A. (2015, May 1). *An evaluation of authorship attribution using random forests*. IEEE Xplore. https://doi.org/10.1109/ICTRC.2015.7156423

Kincaid, J. P., Fishburne, J., Rogers, R. L., & Chissom, B. S. (1975, February 1). *Derivation of New Readability Formulas (Automated Readability Index, Fog Count and Flesch Reading Ease Formula) for Navy Enlisted Personnel*. Apps.dtic.mil. https://apps.dtic.mil/sti/citations/ADA006655

Koppel, M., Schler, J., & Argamon, S. (2011). Authorship attribution in the wild. *Language Resources and Evaluation*, *45*(1), 83–94. JSTOR. https://doi.org/10.2307/41486029

Luyckx, K., & Daelemans, W. (2008). Authorship Attribution and Verification with Many Authors and Limited Data. *ACL Anthology*, 513–520. https://aclanthology.org/C08-1065/

Mardaga, H. (2012). Hapax Legomena: A Neglected Field in Biblical Studies. *Currents in Biblical Research*, *10*(2), 264–274. https://doi.org/10.1177/1476993x11398845

Mercer, D. (2017, October 25). *David L. Wallace, statistician who helped identify Federalist Papers authors, 1928-2017*. University of Chicago News. https://news.uchicago.edu/story/david-l-wallace-statistician-who-helped-identify-federalist-papers-authors-1928-2017

Plotnikov, A. V. (2020). Gunning fog-index measurement of customer reviews of the Russian Agricultural Bank. *IOP Conference Series: Earth and Environmental Science*, *548*(2), 022046. https://doi.org/10.1088/1755-1315/548/2/022046

Rogers, A., Kovaleva, O., & Rumshisky, A. (2020). A Primer in BERTology: What we know about how BERT works. *ArXiv:2002.12327 \[Cs\]*. https://arxiv.org/abs/2002.12327

Rossiello, G., Basile, P., & Semeraro, G. (2017). *Centroid-based Text Summarization through Compositionality of Word Embeddings* (pp. 12–21). Association for Computational Linguistics. https://aclanthology.org/W17-1003.pdf

Schler, J., Koppel, M., Shlomo Argamon, & Pennebaker, J. W. (2005). Effects of Age and Gender on Blogging. *National Conference on Artificial Intelligence*, 199–205.

Sohrab, M. G., Miwa, M., & Sasaki, Y. (2015). Centroid-Means-Embedding: An Approach to Infusing Word Embeddings into Features for Text Classification. *Lecture Notes in Computer Science*, 289–300. https://doi.org/10.1007/978-3-319-18038-0_23

Stamatatos, E. (2009). A survey of modern authorship attribution methods. *Journal of the American Society for Information Science and Technology*, *60*(3), 538–556. https://doi.org/10.1002/asi.21001

Sultana, S. J., Hoque, M. N., Chy, A. N., & Md Hanif Seddiqui. (2024). *Enhanced Hate Speech Detection through Mean-Pooling in Embedding Fusion*. 1833–1838. https://doi.org/10.1109/iccit64611.2024.11021822

Wermer-Colan, A. (2023, November 22). *Research Guides: Stylometry Methods and Practices: Home*. Guides.temple.edu. https://guides.temple.edu/stylometryfordh