📄 Arabic Text Sentiment Analysis: Reinforcing Human-Performed Surveys with Wider Topic Analysis

Summary

The authors map research on Arabic sentiment analysis (ASA), a field with far less work than English sentiment analysis. Arabic is hard for this task: it has rich morphology, often drops diacritics, and mixes Modern Standard Arabic (MSA) with many regional dialects. The review has two layers. First, a manual, PRISMA-style systematic review of 133 English-language ASA papers from 2002–2020, drawn from SAGE, IEEE, Springer, WILEY and Google Scholar. Second, an automatic analysis of 2297 open-access ASA publications from 2010–2020, using topic modelling and temporal trends to test whether the manual themes hold at scale. The main findings:

• SVM, Naive Bayes and hybrid methods are the most common strong performers.
• Deep models such as LSTM can be more accurate but are limited by small corpora.
• Lexicons with wide coverage are MSA-only, and dialect lexicons are small.
• Tweet corpora are only mid-sized, and practical ASA tools for industry are scarce.

The work gives a reference map of methods and resources and points to dialect data as the main gap.

Key points

• Screening funnel: 702 records found (234 from databases, 468 from grey literature), 15 duplicates removed, 554 excluded at title, abstract or method level, and 133 papers kept after full-text review.
• The automatic in-breadth study covers 2297 publications. The best topic model used 201 topics with ɑ = 0.01 and reached coherence 0.39. The resulting topics were then hand-labelled into themes such as Transformer Networks, Convolutional Networks, Recurrent Networks, Bayesian Learning, Semantic Web, Twitter and Facebook.
• The approach families identified are machine learning, lexicon-based and hybrid. Transfer learning was nearly absent: only one study in 2019 and one in 2020.
• Lexicon resources vary widely in size:
  • ArSenL: 157,969 words
  • SLSA: 35,000 lemmas
  • NileULex: 5953 Egyptian dialect words and phrases
  • SauDiSenti: a Saudi dialect sentiment lexicon
  • smaller lexicons that map Jordanian dialect, Arabizi or emoticons onto MSA
• Saudi tweet corpora range from about 1,331 to 17,573 tweets (AraSenti-Tweet), and most are not publicly available. Six dialect groups are distinguished: Gulf, Yemeni, Iraqi, Egyptian, Levantine and Maghrebi.

Interesting ideas

• I1. The authors check a small, carefully hand-read sample against a large automatic analysis of thousands of documents, which turns the hand-read conclusions into claims with quantitative backing. — For Mova (learn): Mova's evaluation could pair a small, manually audited set of parses or frames with corpus-wide automatic statistics, flagging cases where the two views disagree as candidates for repair rules.
• I2. The topic model is deliberately run with very many fine topics, and humans then merge and name the useful ones, so the machine supplies clusters while people supply the labels. — For Mova (world): the same over-cluster-then-curate loop could help grow ontology categories or verb-class groupings from corpus co-occurrence, while keeping each final category explainable and named.
• I3. Frequent multiword units such as naive_bayes or twitter_feed are joined into single tokens before modelling, so they are not split into misleading separate words. — For Mova (tokens): detecting and lemmatizing multiword expressions as one unit before tagging would give later stages cleaner lemmas for names and fixed terms.
• I4. Hierarchical sentiment classification first decides whether a text is subjective or objective, and only then assigns polarity to the subjective part. — For Mova (frames): prag could first decide whether a clause is evaluative or factual before assigning stance or irony, which avoids forcing polarity onto plain statements.
• I5. Several of the surveyed lexicons translate dialect, Arabizi or emoticons into the standard variety, so existing standard-language resources can be reused for informal text. — For Mova (tree): a normalisation layer that maps informal or dialectal English onto standard forms before parsing would let the existing treebank-trained parsers cover social-media text.