Supervised Learning for Negation Scope Detection n Arabic Texts

Abstract

The research in detecting the negation in Arabic is limited due to the unavailability of Arabic corpora targeting this phenomenon. The negation detection affects a set of subfields in the Arabic Natural Language Processing (ANLP), including sentiment analysis and medical information retrieval. Therefore, a corpus is manually annotated with negation for targeting this deficiency in the Modern Standard Arabic (MSA) and Classical Arabic (CA) texts. This corpus is collected from various sources, including King Saud University Corpus of Classical Arabic (KSUCCA) and Wikipedia. It includes texts from various topics, like religion, sports, science, biography, health, technology, education, and history. In addition, we propose a supervised-based learning system for the problem of negation scope detection in Arabic texts. Our system depends on Word2Vec and FastText word embeddings with two different classifiers: the Bidirectional Long Short-Term Memory (BiLSTM) and the Support Vector Machined (SVM) as a baseline system. The results show that one of the FastText-BiLSTM based models achieved a classification accuracy of 93% with F1 score of 89%. The fact that the results of the supervised learning are encouraging further proves the point that the treatment of the negation phenomenon is tractable.