Elamur Research⌕
☰
home / papers / 2017 / Semantic Similarity Analysis for Paraphrase Identification in Arabic Texts

Semantic Similarity Analysis for Paraphrase Identification in Arabic Texts

Paper · 2017 · Adnen Mahmoud and Mounir Zrigui

Summary

The paper addresses paraphrase identification in Arabic, framed as a step in plagiarism detection. The task is hard because Arabic is highly inflected and agglutinative and is usually written without diacritics. One surface form can therefore belong to several lexical categories and carry quite different meanings. The authors build a corpus-based similarity pipeline in four steps:

  • Train word2vec (Skip-gram) embeddings on a large Arabic Wikipedia dump.
  • Build a vector for each sentence by averaging its word vectors, weighted by TF-IDF.
  • Compare every sentence of a suspect document against every sentence of the source document, using both Cosine Similarity and Euclidean Distance, merged into one combined score.
  • Declare a paraphrase when that score passes a tuned threshold.

Evaluation uses the history portion of the Open Source Arabic Corpus (OSAC). The full combination of word2vec, TF-IDF and the merged metrics scores best, reaching roughly 0.85 precision and 0.84 recall. The work matters as an early, simple baseline showing that weighted embeddings help Arabic text-similarity tasks, where resources are scarce.

Key points

  • OSAC contains 22,429 documents in 10 categories. The experiments use a subset of 3,233 history documents.

  • Skip-gram word2vec was trained on more than 350 million words from Wikipedia. The context window is two words on each side of the target, and the paper sets k = 5.

  • Each sentence vector is a TF-IDF-weighted mean of its word vectors. Sentences of a text form the columns of a matrix V.

  • The paraphrase threshold α = 0.3 was found by repeated trials on training data.

  • Table 3 is garbled in the extracted text. Read as an ablation, it shows:

    ConfigurationPrecisionRecall
    word2vec + combined metric≈0.81≈0.79
    TF-IDF + combined metric≈0.83≈0.81
    word2vec + TF-IDF + combined metric≈0.85≈0.84

Interesting ideas

  • I1. Weighting each word's contribution by TF-IDF lets descriptive words dominate a sentence's representation while frequent function words fade, a cheap salience signal usable beyond similarity.
  • I2. Rather than comparing whole documents at once, every suspect sentence is scored against every source sentence, producing a fine-grained alignment profile instead of a single number.
  • I3. Two complementary metrics, an angle-based one and a distance-based one, are fused and then gated by a single threshold tuned on training data. Combining measures with different blind spots can be more robust than either alone.
  • I4. The results table is essentially an ablation that adds components one at a time, which makes each part's contribution visible.
  • I5. The paper's catalogue of Arabic ambiguity, including category ambiguity of one surface form and an attached conjunction letter that may or may not be part of the word, shows how segmentation errors can flip meaning before any semantics happens.

For Mova

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

  • I1 (world): A corpus-level TF-IDF salience score could rank which entities and predicates in the world of a text are central, guiding which facts the answerer checks first.
  • I2 (answer): Aligning each question sentence against each story sentence, but over frames rather than vectors, would give an explainable trace of which passage supports an answer.
  • I3 (learn): When several witness parsers or scorers judge silver data, fusing distinct agreement measures with a held-out-tuned cutoff could make acceptance decisions less brittle.
  • I4 (learn): Evaluations of new repair rules or teacher tiers could routinely report incremental add-one-component rows alongside negative controls.
  • I5 (tokens): A similar inventory of English clitic and category-ambiguity cases ("'s", noun/verb homographs) could drive targeted tagger tests, since errors there propagate into the tree and frames.

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

 Institutions

🎓University of MonastirMonastir · Tunisia6 works · 47 citations🎓Tunis UniversityTunis · Tunisia4 works · 28 citations

 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 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Adnen Mahmoud and Mounir Zrigui. 2017. Semantic Similarity Analysis for Paraphrase Identification in Arabic Texts. Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{mahmoud2017semantic,
  title = {{Semantic Similarity Analysis for Paraphrase Identification in Arabic Texts}},
  author = {Adnen Mahmoud and Mounir Zrigui},
  booktitle = {Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation},
  year = {2017},
  url = {https://aclanthology.org/Y17-1037/}
}

References

24 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abderahman Y. A., Khalid A. and Osman I. M.. (2016). A survey of plagiarism detection for Arabic documents, International Journal of Advanced Computer Technology IJACT, volume 4, n. 6, pp. 34-38. find
  2. Alaa Z., Tiun S. and Abdulameer M. H. (2016). Cross-language plagiarism of Arabic-English documents using linear logistic regressing, Journal of Theoretical and Applied Information Technology, 2005 - 2015 JATIT & LLS, Volume 83. No.1, , pp. 20-33. find
  3. Aliwy A. H. (2012). Tokenization as Preprocessing for Arabic Tagging System, International Journal of Information and Education Technology, volume 2, No. 4, pp. 348-353. find
  4. Bihi A. (2017). Analysis of similarity and differences between articles using semantics, Université Malardalen, Sweden. find
  5. Boudhief A., Maraoui M. and Zrigui M. (2014). Elaboration of a model for an indexed base for teaching Arabic language to disabled people, 6 th International Conference on CIST, pp110-116. find
  6. Gharavi E., Bijari K., Zahirnia K. and Veisi H.. (2015). A deeplearning approach to Persian plagiarism detection, India. find
  7. Hafeez S. and Patil B. (2017). Using Explicit Semantic Similarity for an Improved Web Explorer with ontology and TF-IDF, International Journal of Advance Scientific Research and Engeneering Trends, Volume 2, Issue 7, pp. 171-173. find
  8. Hussein A. S. 2017. Arabic document similarity analysis using Ngram and Singular Value Decomposition, 9 th International Conference on Research Challenges in Information Science RC IS. find
  9. Kenter T. and Rijke M. D. (2015). Short text similarity with word embeddings, International Conference on Information and Knowledge Management CIKM’15, Australia. find
  10. Law J., Zhuo H. H., He J. and Rong E. (2017). LTSG: Latent Topical Skip-Gram for Mutually Learning Topic Model and Vector Representations, Cornell University Library, Coputer Science, Computation and language, United States. find
  11. Lihioui C., Zouaghi A. and Zrigui M. (2014). Towards a hybrid approach to semantic analysis of spontaneous Arabic speech, International Journal of Computational Linguistics and Applications, volume 5, n. 2, pp. 165-193. find
  12. Meddeb O., Maraoui M. and AlJawerneh S. (2016). Hybrid modeling of an offline arabic handwriting recognition system AHRS, International Conference on Engineering & MIS ICEMIS, Maroc Mikolov T., Sustskever I., Chen K., Corrado G., and Dean J. (2013). Distributed representations of words and phrases and their compositionality, Neural Information Processing Systems NIPS, United States Vo N. P. A., Magnolini S. and Popescu O. (2015). Paraphrase identification and semantic similarity in Twitter with simple features, Proceedings of Social NLP ,pp. 10-19, Colorado.
  13. Negre E. (2013). Comparaison de textes: quelques approche , Cahier du LAMSADE 338, Laboratoire d’Analyses et Modélisation de Systèmes pour l’Aide à la Décision UMR 7243, Paris. find
  14. Niraula N. B., Gautam D., Banjadae R., N. Maharjan, and Rus V. (2015). Combining word representations for measuring word relatedness and similarity, Twenty Eight International Florida Artificial Intelligence Research Society Conference, Florida Prazak O., Steinberger D., Konopik M., and Brychain T. (2012). Interpretable semantic textual similarity with distributional semantics for chunks, Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp1212-1222. find
  15. Peng X. and Gildea D. (2016). Exploring phrase-compositionality in skip-gram models, Cornell University Library, Computer Science, Computation and language, United States. find
  16. Samuel D. S. (2016). On the use of vector representation for improved accuracy and currency of Twitter POS Tagging, Dalhousie University, Halifax, Nova Scotia. find
  17. Shenoy M. K., Set D. C. and Achrya U. D. (2012). Semantic plagiarism detection system using ontology mapping, Advanced Computing: An International Journal ACIJ, volume 3, n. 3, pp. 59-62. find
  18. Shuka V., Khan F. and Mody K. (2016). Plagiarism detection for document, International Journal on Recent and Innovation Trends in Computing and Communication, volume 4, issue 2, pp. 175-178. find
  19. Sindhu L., and Idicula S. M. (November 2015). SRL based plagiarism detection system for Malayalam documents, International Journal of Computer Science Issues IJCSI, volume 12, issue 6, pp. 91-97. find
  20. Towne W. B., Rosé C. P. and Herbsleb J. D. (2016). Measuring similarity: LDA and Human Perception, ACM Transaction on Intelligent Systems and Technology, pp. 1-29, volume 7, n. 2. find
  21. Yih W. T., Zweig G. and Platt J. C. (2012). Polarity inducing Latent Semantic Analysis, Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1212–1222, Korea. find
  22. Zouaghi A., Zrigui M., Ben Ahmed M. and Antoniadis G. (2007). L’influence du contexte sur la compréhension de la parole spontanée, Proceedings de la Conference Traitement Automatique de La Langue Naturelle TALN'07 Zouaghi A., Zrigui M. and Antoniadis G. (2008). Compréhension automatique de la parole arabe spontanée , Traitement Automatique des Langues, Beljique. find
  23. Zouaghi A., Marhbène L. and Zrigui M. (2012). A hybrid approach for Arabic word sense disanbiguation, Internatonal Journal of Computer Processing of Languages, volume 24, n.2, pp. 133-151. find
  24. Zrigui S., Zouaghi A., Ayadi R., Zrigui M. and Zrigui S. (2016). ISAO: An intellgent system of opinion analysis, Research in Computing 110 , pp. 21-31. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mahmoud-2017-semantic-similarity-analysis-paraphrase-identification/