Arabic Paraphrasing Detection Using Multiple Extracted Features

Paraphrase detection identifies whether two given texts have the same meaning despite differences in wording and structure. Improving these applications' accuracy and effectiveness is crucial, especially for Arabic language, where additional challenges exist due to its complex nature and limited online resources. This paper investigates and compares the applicability of various statistical techniques and machine learning classifiers for identifying paraphrasing in Arabic texts. Given the limited availability of Arabic datasets, A new dataset is proposed that comprises 143 short texts collected from Al-Manara journal, which was paraphrased under the guidance of an Arabic linguistics specialist. This study examines the effectiveness of various statistical and machine learning techniques for feature extraction, such as unigram, bigram, TF-IDF, and word embedding. Moreover, it assesses the performance of several machine learning classifiers for paraphrase detection, including SVM, DT, K-NN, RF, and deep learning (BERT) using different metrics such as accuracy, precision, recall and F-measure. Experiments showed that bigram technique performed the best. Additionally, all machine learning classifiers outperformed the bigram technique in all metrics.