💻 computer science

Improving Arabic Text Sentiment Analysis using Aspect-based

This study proposes a hierarchical network with word-level attention and ratio-based weighting for Arabic aspect-based sentiment analysis, demonstrating that using CamelTools embeddings significantly outperforms Word2Vec across multiple datasets in terms of accuracy and F1-score.

Original authors: Faisal Mehmood, Touqeer Abbas, Sami Ullah

Published 2026-09-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Faisal Mehmood, Touqeer Abbas, Sami Ullah

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, millions of people turn to social media to share their thoughts, complaints, and praise. These digital conversations cover everything from politics to the latest smartphone, creating a vast ocean of unstructured text that reflects how people truly feel. For businesses and political leaders, understanding these feelings is crucial; knowing whether the public is happy or angry can shape marketing strategies, improve products, and guide policy decisions. However, reading through this endless stream of posts is impossible for humans, and teaching computers to understand the nuances of language is a difficult task. This challenge is even greater with Arabic, a language rich in dialects and complex grammar, where the same word can carry different weights of meaning depending on the context. While computers have become quite good at reading English text, they often struggle to grasp the subtle differences in Arabic sentiment, particularly when a person is talking about a specific part of a product or service rather than the whole thing.

Researchers Faisal Mehmood, Touqeer Abbas, and Sami Ullah set out to solve this specific problem by building a new kind of computer model designed to read Arabic tweets with greater precision. Instead of treating a sentence as a single block of text, their approach focuses on "aspect-based" sentiment. This means the model is designed to determine the feeling toward a specific topic, but it requires that the specific topic (or "aspect") be provided to it beforehand. For example, if a review says, "The battery life is terrible but the screen is beautiful," the model does not automatically decide which part is being discussed; instead, it takes the sentence and a specific aspect (like "battery") as input to determine if the sentiment toward that specific aspect is negative or positive. To achieve this, the team created a system that pays special attention to the most important words in a sentence, ignoring the less significant ones that do not carry the emotional weight.

The researchers tested their idea using three different collections of Arabic tweets, which included posts from Egypt, Jordan, and a mix of general topics. Before the computer could learn, the team had to clean the data, removing things like links, repeated letters, and common words that do not add meaning, such as "the" or "of." They then fed the cleaned text into a neural network, a type of computer program inspired by the human brain, which is particularly good at recognizing patterns in sequences. The core of their innovation was a layer of "attention" that acts like a spotlight. As the computer reads a sentence, this spotlight highlights the words that are most relevant to the specific aspect being analyzed, allowing the model to weigh them more heavily than the surrounding words. They compared their new method against older techniques that treated all words equally and against other advanced models that had been used previously.

The results showed that their approach was significantly more accurate. When tested on the different datasets, the new model consistently outperformed the previous best methods. Specifically, the study found that using the CAMeL Tools preprocessing method yielded better results than using Word2Vec embeddings. On the ArTwitter dataset, this preprocessing improvement resulted in a 4.50% increase in accuracy and a 4.70% increase in the F1-score. On the ASTD dataset of Egyptian tweets, the use of CAMeL Tools over Word2Vec led to a 2.60% improvement in accuracy and a 2.44% increase in the F1-score. On the AJGT dataset, the improvement was 2.10% in accuracy and 3.34% in the F1-score. The researchers found that using a specific tool called CAMeL Tools to process the Arabic text worked better than older methods of converting words into numbers. By focusing on the importance of individual words within a sentence, the model was able to capture the subtle shifts in sentiment that earlier systems missed.

This work suggests that for computers to truly understand human opinion in Arabic, they must be taught to look at the structure of a sentence and recognize which words carry the most emotional weight. The study does not claim to have solved every problem in language processing, but it demonstrates that a targeted approach, one that isolates specific topics and weighs words differently, leads to better results. The researchers hope that by refining how machines analyze these linguistic nuances, they can help organizations make better sense of the massive amount of feedback generated on social media every day. The findings indicate that with the right tools, the complexity of Arabic dialects and grammar can be navigated effectively, turning raw text into clear, actionable insights.

Technical Summary: Improving Arabic Text Sentiment Analysis using Aspect-based Sentiment and Word-level Hierarchical Attention Networks

Problem Statement
Sentiment analysis (SA) is a critical tool for interpreting user opinions on social media for political and corporate decision-making. While deep learning has advanced English SA, Arabic SA faces unique challenges due to the language's informal, unstructured nature and complex linguistic nuances. Specifically, traditional sentence-level SA often fails to capture sentiment toward specific aspects (e.g., a specific feature of a product) within a sentence. Furthermore, existing neural models often overlook the varying degrees of importance individual words carry regarding a specific aspect, treating all words in a sequence with equal weight. This paper addresses the need for a model that can effectively perform Aspect-Based Sentiment Classification (ABSC) in Arabic by identifying the relative significance of words within a sentence.

Methodology
The authors propose a hierarchical neural network architecture that integrates word-level attention mechanisms with Bidirectional Long Short-Term Memory (biLSTM) layers. The system operates through the following stages:

Data Preprocessing: The study utilizes three benchmark Arabic datasets: ASTD (Egyptian tweets), AJGT (Jordanian dialect and MSA), and ArTwitter. Data cleaning involves removing noise (URLs, hashtags, symbols), tokenization, and normalization (removing diacritics, standardizing characters, and removing stop words). The authors employ CAMeL Tools, a Python-based open-source NLP suite, for preprocessing and dialect identification, citing its superiority over other tools for Arabic text.
Word Representation: Tweets are converted into fixed-length vectors using AraVec (Word2Vec skip-gram model trained on Twitter data) with 300 dimensions.
Hierarchical Architecture:
- Word Layer Encoding: The model creates aspect-augmented embeddings by concatenating word embeddings with position embeddings and the aspect representation vector.
- Bi-LSTM Encoding: Forward and backward biLSTM layers process these embeddings to capture contextual information from both past and future words within a clause.
- Word-Level Attention: A critical component of the proposed model is the attention layer. It calculates attention weights ( $\alpha_{ij}$ ) for each word based on its relevance to the specific aspect. This allows the model to focus on informative words (e.g., adjectives or specific sentiment carriers) while downplaying less relevant terms.
- Classification: The weighted sum of word annotations forms a clause vector, which is passed through a SoftMax layer to predict the sentiment polarity (positive, negative, or neutral).
Training: The model is trained using a cross-entropy loss function with 10-fold cross-validation. Hyperparameters include a learning rate of 0.001, a dropout rate of 0.25, and L2 regularization.

Key Contributions
The paper outlines four primary contributions:

Novel Architecture: Introduction of a new aspect-based SA model that leverages a word-level attention mechanism to focus specifically on important words within text relative to a target aspect.
Performance Improvement: Demonstration of significant improvements in accuracy and F1-scores over state-of-the-art approaches across three distinct Arabic sentiment datasets.
Comprehensive Evaluation: Extensive benchmarking against recent deep learning models (including CNNs, LSTMs, and hybrid models) and traditional machine learning classifiers (SVM, Naive Bayes, KNN, Logistic Regression).
Dialect Robustness: Application of the model across various Arabic dialects (Egyptian, Jordanian, and Modern Standard Arabic) to ensure effectiveness in an unevenly characterized linguistic landscape.

Experimental Results
The proposed "hybrid" model (biLSTM with attention) was evaluated against existing baselines on the ASTD, AJGT, and ArTwitter datasets.

ArTwitter Dataset: The proposed model achieved an F1-score of 95.93% and accuracy of 95.83%, outperforming the previous best biLSTM model (92.39% F1) and hybrid CNN-LSTM models.
ASTD Dataset: The model reached an F1-score of 97.13% and accuracy of 97.43%, significantly surpassing the CNN baseline (82.48% F1) and the previous biLSTM baseline (79.25% F1).
AJGT Dataset: The model achieved an F1-score of 96.43% and accuracy of 95.63%, outperforming Logistic Regression (83.96% F1) and other traditional classifiers.

A comparative analysis highlighted that using CAMeL Tools for preprocessing yielded better results than standard Word2Vec approaches, showing average improvements of 2.60% in accuracy and 2.44% in F1-score on the ASTD dataset. Visualization of the attention weights confirmed that the model successfully identified and weighted sentiment-carrying words (e.g., "only," "love") more heavily than neutral terms.

Significance and Claims
The authors claim that their work provides a systematic evaluation of Arabic ABSA literature and overcomes inadequacies in previous reviews by introducing a framework that explicitly accounts for word-level importance. The study asserts that by integrating word-level attention with biLSTM, the model captures the nuanced semantic and syntactic relationships inherent in Arabic text more effectively than models that treat sentences as uniform sequences. The authors conclude that their approach offers a superior method for learning semantic information in Arabic sentiment analysis, particularly in handling the diverse importance levels of words associated with specific aspects. The paper modestly notes that while the model is effective, future work could further optimize performance by incorporating discourse analysis between consecutive sentences.

More like this