Enhancing Arabic Word Stemming Using the Char Stemmer Deep Learning Model

نویسندگان
قائم مقامی صفا ID 1
هاشمی کرویی محمدمهدی ID 2

1 دانشگاه آزاد اسلامی

2 دانشگاه علامه طباطبایی

doi
10.37264/JIQS.V5I1.6
چکیده

The stemming of Arabic words is a crucial step in several text processing tasks, including text mining, information retrieval, and natural language processing. Arabic stemmers face many challenges, mainly due to the complex nature of Arabic words and their different writing styles. To address these challenges, this paper presents a novel approach integrated into a Python module for improving conventional Arabic stemming. The proposed approach employs neural network architectures to achieve high accuracy in Arabic text stemming. We utilize a Bi-LSTM architecture to extract Arabic stems and construct sequence-to-sequence stemming models. The proposed model is named Char Stemmer. To evaluate this model, we use a benchmark dataset of Qur’anic words and classical Arabic texts. Extensive experiments were conducted to evaluate the model, including comparisons with several well-established Arabic stemmers, namely Khoja, Light-10, P-Stemmer, Tashaphyne, and Al-Khalil. The experimental results show that the proposed model achieves a stem accuracy of 93.88%, substantially outperforming conventional rule-based and light stemming methods. The results demonstrate the effectiveness of deep learning sequence models for Arabic stemming and highlight the potential of character-level models for capturing complex morphological structures. The proposed model is generalizable, scales well, and can serve as a domain-independent solution for a wide range of Arabic NLP applications. Furthermore, the model can be directly integrated into Arabic information systems for tasks such as automated text normalization, intelligent search, and knowledge extraction in applied computing environments.

کلیدواژه‌ها
Character-level Seq2Seq model for Classical Arabic stemming: This paper presents a character-level seq2seq stemming model based on Bidirectional LSTM (BiLSTM) as an encoder and LSTM as a decoder. The model is specifically applied to classical Arabic texts, which are known for their rich and complex morphological structures., Application to Qur’anic and Classical Arabic Morphology: Unlike many previous studies that focus on Modern Standard Arabic, this work evaluates the model on classical Arabic data, including Qur’anic text, providing insights into handling traditional morphological patterns., Unified Preprocessing and Evaluation Framework: The proposed model applies a consistent normalization and preprocessing pipeline across all experiments, ensuring a fair comparison between the proposed model and baseline stemmers., Comparative Evaluation with Widely Used Arabic Stemmers: The model is evaluated against several well-known Arabic stemming approaches, including Khoja, Light-10, P-Stemmer, Tashaphyne, and Al-Khalil, using the same dataset and evaluation metrics., Alphabet Normalization: Converts all Arabic characters into a consistent form, addressing the issue of character variability across different Arabic scripts., Digit Normalization: Standardizes Arabic digits to ensure numerical data is uniformly represented., Punctuation Handling: Removes or standardizes punctuation marks, which are generally not required for the stemming process., Diacritic Removal: Deletes diacritical marks that are extensively used in Arabic for disambiguation but are often omitted in everyday typing and text messaging and are therefore not crucial for stemming., Whitespace Management: Eliminates extra spaces to ensure consistent separation between words., Accuracy: Measures the percentage of correct predictions over all predictions made., Precision, Recall, and F1-score: These metrics provide deeper insight into the model's performance, especially in handling class imbalances that are common in character-based tasks., Robust Preprocessing: The model's success begins with comprehensive preprocessing steps that normalize and prepare the classical Arabic texts for machine learning., Character-Level Embedding: By transforming each character into a dense vector representation, the model captures subtle semantic and syntactic nuances of the Arabic language., Bidirectional LSTM Encoder: The utilization of a bidirectional LSTM allows the model to incorporate context from both the beginning and the end of a word when making predictions about its stem., Sequential Decoding: The LSTM decoder predicts one character at a time, using the previous character's context to influence the next., Focused Loss Function: Employing Cross-Entropy Loss ensures that the training process focuses on improving the probabilities of correctly predicted characters., Expanded Training Data: Including a more diverse set of training examples, especially those that represent edge cases, could help improve the model's ability to generalize across less frequent morphological patterns., Hyperparameter Optimization: Further tuning of parameters such as the embedding size, the depth of LSTM layers, or the learning rate could yield improvements in model accuracy and recall., Khoja Stemmer:The Khoja stemmer removes the longest suffixes and prefixes. It then matches the remaining part with verb and noun patterns to derive the stem. It uses several linguistic data files, including lists of characters, punctuation marks, diacritics, and stop words. Khoja has two dictionaries, one for stems and another for roots. The Khoja stemmer is available as a Java implementation (Khoja & Garside 1999)., Light-10:Light-10 performs the following steps for stemming:, It divides the text into words., Normalization converts all (أ, إ, آ) to(ا), (ى) to (ی), and (ة) to (ه)., It removes the letter (و) when used as the conjunction "and"., It removes prefixes and suffixes from words. The prefixes used by Light-10 include (لل, و, فال, کال, بال, وال, ال), while the suffixes include (ی, ة, ه, یة, یه, ین, ون, ات, ان, ها). The Light-10 implementation is available in Java (Larkey et al. 2007)., Tashaphyne:The Tashaphyne stemmer normalizes words and removes diacritics and elongation from input words. It then segments and stems the input using an affix lookup list for different levels of stemming. The Tashaphyne implementation is available in Python (Al-Khatib et al. 2023)., NLTK (Natural Language Toolkit):NLTK is a collection of open-source programs for NLP. It includes over 50 corpora and lexical resources for tasks such as stemming, tokenization, classification, and parsing. NLTK was developed by Steven Bird and Edward Loper (Elbes et al. 2019)., P-Stemmer:P-Stemmer, created by Kanan and Fox, is an updated version of Larkey’s light stemmer. It removes only the prefixes rather than removing both prefixes and suffixes, resulting in effective and reliable stemming performance. The P-Stemmer implementation is available in Python (Kanan et al. 2019)., Superior Accuracy: The proposed model achieves significantly higher accuracy (93.88%) compared to other Arabic stemmers., Precision-Recall Trade-off: While the proposed model achieves the highest precision (82.01%), it shows a lower recall (70.46%) compared to the P-Stemmer (83%). This behavior reflects a precision-oriented strategy, which is preferable in downstream NLP tasks where over-stemming can introduce semantic errors., F1-Score: The proposed model achieves an F1-score of 75.28%, ranking second after P-Stemmer (77%), demonstrating balanced performance across precision and recall. Figure 4 illustrates the comparative performance of the proposed model and the other Arabic stemmers., RIS, EndNote, Mendeley, BibTeX, APA, MLA, HARVARD, CHICAGO, VANCOUVER