Intrinsic Plagiarism Detection in Armenian Texts Using Stylometric Analysis

This repository contains the models and resources for our stylometry-based intrinsic plagiarism detection research for the Armenian language. We provide the evaluation script, as well as the datasets and a set of manually compiled stylometric feature extraction resources that were used for evaluation.

Tasks and models

Intrinsic stylometric analysis was studied from the perspective of two PAN @ CLEF shared tasks: Style Change Detection and Style Breach Detection. The first task is about detecting whether there is a style change within the document. For this task, we adapted and studied the clustering-based Nath et al. model and Zlatkova et al. ensemble of classifiers:

• Nath et al. are using threshold clustering on text windows' vectors. Style change is detected if the number of obtained clusters is greater than one.

• Zlatkova et al. are using a weighted combination of 4 classifiers(SWM, Random Forest, AdaBoost and Multi-layer Perceptron), applied to the feature vectors. Then the obtained results are analysed by logistic regression to detect a style change.

The second task is about detecting the exact indices where style change occurs in the text. For this task, we studied the model from Karas et al. and an agglomerative clustering-based new approach, inspired by the top solutions for the Authorship Clustering task. For both models, we extract and use our own stylometric features set.

• Karas et al. model represents a statistic approach using Wilcoxon test to compare neighboring paragraphs' similarity score. Paragraph pairs with the least 30% of test scores are considered to have style differences and a style breach between them is fixed.

• Agglomerative Clustering model is using clustering approach to group the paragraphs by their clusters and indicates a style breach if there is a cluster change between two neighboring paragraphs.

Dataset

The dataset was generated automatically by combining fragments from different texts on the same topic. We generated multiple test sets to cover various writing genres:

1. PhD theses, collected from the official website.
2. Encyclopedic articles, collected from Wikipedia and Armenian encyclopedia.
3. Academic books, collected from history textbooks for high school 7-9 grades and for universities.
4. Fiction, collected from "Pheasant" by Axel Bakunts and the its movie adaptation by Hrant Matevosyan.
5. News, collected from NewsHub news aggregator.

Stylometric features

We used a combination of character-, word-, sentence- and paragraph-level features, implemented in feature_extraction module:

a) Character level: features describing the usage of suffixes, prefixes and punctuation.

b) Word level: this group includes features based on the usage of foreign, informal words, specific n-grams, preferred style of writing abbreviations and numerals.

c) Sentence level features consist of morphological and syntactic features as well as general sentence length features.

d) Paragraph level readability features: Flesch reading ease, SMOG grade, Flesch-Kincaid grade, Coleman-Liau index, automated readability index, Dale-Chall readability score, difficult words, Linsear write formula and Gunning fog.

For more details about the models and dataset construction, refer to the paper.