Elamur Research⌕
☰
home / papers / 2015 / Rational Kernels for Arabic Stemming and Text Classification

Rational Kernels for Arabic Stemming and Text Classification

Paper · 2015 · Attia Nehar, Djelloul Ziadi and Hadda Cherroun

Summary

The paper addresses two linked tasks: reducing Arabic words to their roots, and sorting Arabic documents into topic categories. The authors build a stemmer that needs no root dictionary. It encodes Arabic morphological patterns (measures), prefixes and suffixes as weighted finite-state transducers. These are combined through union and concatenation, and the result is composed with a word's transducer to produce candidate three-letter roots. The best candidate is then chosen using statistics on how often each character occurs in Arabic roots.

Because stemming is done with transducers, each document becomes a transducer as well. Rational kernels, specifically n-gram kernels in OpenKernel with LibSVM, then compare documents directly.

On three hand-annotated word sets, the stemmer is more accurate than the dictionary-free Al-Serhan stemmer but less accurate than the dictionary-based Khoja stemmer. On the Saudi Press Agency corpus, stemming improves accuracy, recall and F1, while precision drops slightly. The 3-gram kernel performs best. This shows that a fixed, linguistically motivated finite-state model can compete without maintaining a lexicon.

Key points

  • Stemmer evaluation used three gold sets annotated with roots and checked by Arabic language experts:
    • Gold1: 679 words sampled from the Corpus of Contemporary Arabic.
    • Gold2: 844 words, built in-house.
    • Gold3: 1,000 words, built in-house.
  • Average stemming accuracy was 85.30% for Khoja, 75.60% for the proposed stemmer and 52.48% for Al-Serhan. The new stemmer scored best on Gold3 (80.30%).
  • Classification used the SPA dataset: 1,526 documents in six categories (culture, economics, social, politics, general, sports), split 80/20 into 1,214 training and 312 test texts.
  • Three setups were compared: no stemming, the Al-Serhan stemmer, and the new stemmer, each with 2-, 3- and 4-gram kernels.
    • The 3-gram kernel gave the best accuracy, recall and F1. The authors link this to most Arabic words having three-letter roots.
    • Sport was the easiest class and General the hardest.
  • A toy example shows the effect of preprocessing on document similarity: the 3-gram Euclidean distance between two related sentences fell from 0.25 to 0.18 after stemming.

Interesting ideas

  • I1. Morphology is modelled as separate small transducers for prefixes, patterns and suffixes. These are concatenated, noun and verb branches are unioned, and the result is composed with the input word, so every analysis corresponds to a traceable path.
  • I2. The analyser first enumerates every candidate stem allowed by the patterns and only then picks one using a lightweight character-statistics prior, keeping hard rules apart from soft ranking.
  • I3. Documents are turned into automata so that similarity becomes a rational kernel computed by transducer composition, with no hand-built feature vectors.
  • I4. The ablation holds the classifier fixed and swaps only the preprocessing (none, a baseline stemmer, the new stemmer). It reports precision separately, which exposes a small loss that the other metrics hide.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tokens): crate en could express English affix and lemma rules as composable finite-state pieces, so each lemma comes with the exact rule path that produced it, which keeps the step explainable.
  • I2 (tokens): ambiguous English lemmas and tags (e.g. -ing nouns vs. verbs) could be produced by rules as a candidate set, with a separate, inspectable scorer that chooses among them.
  • I3 (back): the text → graph → text compressor could score round-trip fidelity with an n-gram kernel over lemma sequences, giving an automata-based closeness measure between the source and the regenerated text.
  • I4 (learn): evaluations of new morphology or repair rules should include a "rule off" control and per-metric, per-class breakdowns, so regressions on one measure are not hidden by an averaged gain.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksMorphology · 405 works

Cite

Cite as Attia Nehar, Djelloul Ziadi and Hadda Cherroun. 2015. Rational Kernels for Arabic Stemming and Text Classification. arXiv:1502.07504.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1502.07504 · arXiv 1502.07504

@misc{nehar2015rational,
  title = {{Rational Kernels for Arabic Stemming and Text Classification}},
  author = {Attia Nehar, Djelloul Ziadi and Hadda Cherroun},
  year = {2015},
  doi = {10.48550/arxiv.1502.07504},
  eprint = {1502.07504},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1502.07504}
}

References

23 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Al-Nashashibi, M., Neagu, D., Yaghi, A.: Stemming techniques for Arabic words: A comparative study. In: Computer Technology and Development (ICCTD),, pp. 270 –276 (2010) find
  2. Al-Serhan, H., Shalabi, R.A., Kannan, G.: New Approach For Extracting Arabic Roots. In: Proceedings of The 2003 Arab Conf. on Infor. Technology, pp. 42–59. Alexandria, Egypt (2003) find
  3. Aljlayl, M., Frieder, O.: On Arabic Search: Improving the Retrieval Effectiveness Via Light Stemming Approach. In: ACM Eleventh Conference on Infor. and Knowledge Management, pp. 340–347 (2002) find
  4. Allauzen, C., Riley, M., Schalkwyk, J., Skut, W., Mohri, M.: OpenFst: A General and Efficient Weighted Finite-State Transducer Library. In: Proceedings of the Ninth International Conference on Implementation and Application of Automata, (CIAA 2007), LNCS , vol. 4783, pp. 11–23. Springer (2007). http://www.openfst.org ↗
  5. Alsaleem, S.: Automated Arabic Text Categorization Using SVM and NB. Int. Arab J. e-Technol. 2 (2), 124–128 (2011) find
  6. Althubaity, A., Almuhareb, A., Alharbi, S., Al-Rajeh, A., Khorsheed, M.: KACST Arabic Text Classification Project: Overview and Preliminary Results. In: Proceedings of The 9th IBIMA conference on Information Management in Modern Organizations (2008) find
  7. Berstel, J.: Transductions and Context-Free Languages. Teubner Studienbücher, Stuttgart (1979) find
  8. Chang, C.C., Lin, C.J.: LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology 2 , 1–27 (2011) find
  9. Cortes, C., Haffner, P., Mohri, M.: Rational Kernels: Theory and Algorithms. J. Mach. Learn. Res. 5 , 1035–1062 (2004) find
  10. Cortes, C., Kontorovich, L., Mohri, M.: Learning languages with rational kernels. In: Proceedings of the 20th annual conference on Learning theory, COLT’07, pp. 349– 364. Springer-Verlag, Berlin (2007) find
  11. Duwairi, R.M.: Arabic Text Categorization. Int. Arab J. Inf. Technol. 4 (2), 125–132 (2007) find
  12. El Kourdi, M., Bensaid, A., Rachidi, T.e.: Automatic Arabic Document Categorization Based on the Naive Bayes Algorithm. In: Proceedings of the Workshop on Computational Approaches to Arabic Script-based Languages, Semitic ’04, pp. 51–58. Association for Computational Linguistics (2004) — Automatic Arabic Document Categorization Based on the Naïve Bayes Algorithm
  13. Gharib, T., Habib, M., Fayed, Z.: Arabic Text Classification Using Support Vector Machines. International Journal of Computers and Their Applications 16 (4), 192–199 (2009) find
  14. Hadi, W., Thabtah, F., ALHawari, S., Ababneh, J.: Naive Bayesian and k-nearest Neighbour to Categorize Arabic Text Data. In: Proceedings of the European Simulation and Modelling Conference. Le Havre, France, pp. 196–200 (2008) find
  15. Kanaan, G., Al-Shalabi, R., Ghwanmeh, S., Al-Ma’adeed, H.: A Comparison of Text-classification Techniques Applied to Arabic Text. J. Am. Soc. Inf. Sci. Technol. 60 (9), 1836–1844 (2009) find
  16. Khoja, S., Garside, R.: Stemming arabic text. Tech. rep., Computing Department, Lancaster University (1999) find
  17. Khreisat, L.: A machine learning approach for Arabic text classification using N-gram frequency statistics. Journal of Informatrics 3 (1), 72–77 (2009) find
  18. Mesleh, A.: Support Vector Machines based Arabic Language Text Classification System: Feature Selection Comparative Study. In: T. Sobh (ed.) Advances in Computer and Information Sciences and Engineering, pp. 11–16. Springer Netherlands (2008) find
  19. Nehar, A., Ziadi, D., Cherroun, H.: Rational kernels for arabic text classification. In: A.H. Dediu, C. Martı́n-Vide, R. Mitkov, B. Truthe (eds.) SLSP, Lecture Notes in Computer Science , vol. 7978, pp. 176–187. Springer (2013) find
  20. Nehar, A., Ziadi, D., Cherroun, H., Guellouma, Y.: An Efficient Stemming for Arabic Text Classification. In: International Conference on Innovations in Information Technology (IIT), pp. 328 –332 (2012) find
  21. Sawalha, M.: Open-source Resources and Standards for Arabic Word Structure Analysis. PhD, University of Leeds, Leeds (2011) find
  22. Sebastiani, F., Ricerche, C.N.D.: Machine Learning in Automated Text Categorization. ACM Computing Surveys 34 , 1–47 (2002)
  23. Syiam, M., Fayed, Z., Habib, M.: An Intelligent System For Arabic Text Categorization. International Journal of Intelligent Computing and Information Sciences 6 (1), 1–19 (2006) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nehar-2015-rational-kernels-arabic-stemming-text/