--- title: "UWB at Arabic Reverse Dictionary shared task: Computing the" description: "The paper describes a University of West Bohemia entry to the KSAA-RD reverse dictionary shared task. 14 references, cited by 1 paper here." authors: ["Stephen Taylor"] year: 2023 venue: "Proceedings of ArabicNLP 2023" source: "https://aclanthology.org/2023.arabicnlp-1.40/" doi: "10.18653/v1/2023.arabicnlp-1.40" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/taylor-2023-uwb-arabic-reverse-dictionary-shared/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/taylor-2023-uwb-arabic-reverse-dictionary-shared/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/14/) / UWB at Arabic Reverse Dictionary shared task: Computing the meaning of a gloss # UWB at Arabic Reverse Dictionary shared task: Computing the meaning of a gloss **Authors:** [Stephen Taylor](https://md.research.elamur.ai/free/en/authors/stephen-taylor/) · **Year:** 2023 · **Venue:** [Proceedings of ArabicNLP 2023](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2023](https://md.research.elamur.ai/free/en/venues/arabicnlp/2023/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2023.arabicnlp-1.40/) · [PDF](https://aclanthology.org/2023.arabicnlp-1.40.pdf) · [DOI 10.18653/v1/2023.arabicnlp-1.40](https://doi.org/10.18653/v1/2023.arabicnlp-1.40) · [OpenAlex](https://openalex.org/W4389518317) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/taylor-2023-uwb-arabic-reverse-dictionary-shared/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper describes a University of West Bohemia entry to the KSAA-RD reverse dictionary shared task. Given a short Arabic or English definition (gloss), a system must predict the embedding of the defined word sense, either as an SGNS vector or as an Electra vector. The author deliberately avoided neural networks. There were fewer than 50k training examples, and pretrained transformers would bring in outside data. The system works as follows: - Each gloss word is mapped to the sense-IDs it could stand for, using the training dictionary plus a lookup table of normalized and substitute forms called "swords". - One sense is picked per word so that the chosen senses are as mutually close as possible. A randomized greedy search does this. - The chosen senses' vectors are combined as a weighted, normalized sum. - English glosses reach the Arabic space through a cross-lingual transform. The method beat the baseline only on Arabic glosses with SGNS vectors, and ranked low elsewhere. Its value is as a transparent, compositional baseline. It also shows clearly that vocabulary coverage is the bottleneck: about half of gloss words remained unmatched. ## Key points - The data comes from the task organizers: - ar.train.json has 45,200 records, 45,058 sense IDs and 38,498 words, drawn from the LMF Contemporary Arabic Dictionary. - ar.en.train.json has 2,862 records and was used only to fit the cross-lingual mapping. - en.complete.with.id.json has 63,596 records and comes from SemEval-2022 Task 1. - The swords table contains three kinds of alias: - Arabic forms stripped of vowels, with alifs unified and final yaa written as alif maqsura (following Zahran et al.). - Generated inflections of one-word verbs. - Single-word glosses, treated as near-synonyms of the word they define. - maxids() is the sense-selection routine and the main processing bottleneck: - It starts from a random combination of candidate senses and changes one ID at a time, greedily lowering the summed pairwise angles. - It runs up to 100 restarts and keeps the best result. - Vectors forming larger angles with the rest receive smaller weights. - On Subtask 1 SGNS, mean cosine was 0.375 against a baseline of 0.262. SGNS results were consistently better than Electra results. A likely reason is that SGNS gives every sense of a word the same vector, so finding the right word is enough even when the sense is wrong. - The author also tried a percentile-based "rank-similarity" measure instead of angles: - It stores 100 cached cosines per vector, about a third of the size of the vector table. - Building the tables took about 17 minutes on a laptop. - It scored slightly worse than angles. - The measure is asymmetric, and its linear interpolation is least accurate at the extremes. ## Interesting ideas - Choose word senses jointly by finding the most mutually coherent combination, using random-restart greedy search instead of exhaustive enumeration; this scales to long phrases cheaply and stays interpretable. - Aggregate dissimilarity as summed angles rather than summed cosines, because angles change most in the near-neighbour region where distinctions matter, while cosines flatten as they approach 1. - Build an alias layer of normalized spellings, generated inflections, and one-word definitions treated as synonyms that points back into the core lexicon, so that unknown surface forms still reach a known entry. - Measure coverage explicitly (here, about 50% of gloss words were dropped) before blaming the composition method; this shows where the real error lies. - Precompute a compact per-item percentile table of similarity values, turning raw cosines into a rank-aware score without storing the full pairwise matrix. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (frames): Pick VerbNet/WordNet senses for all content words of a clause together by maximizing pairwise compatibility, and log the chosen tuple and its scores as the explanation. - **I2** (world): Use angular distance when ranking ontology or knowledge-graph neighbours so that close candidates are separated more sharply. - **I3** (tokens): Add an explicit alias table that maps unseen inflections and spelling variants to known lemmas, recording which rule produced each alias for traceability. - **I4** (learn): Report the share of tokens or frames that fell out of the lexicon as a standard diagnostic alongside accuracy in every evaluation run. - **I5** (world): Calibrate similarity thresholds per concept in the compiled knowledge so that "close" means the same thing for dense and sparse regions of the embedding space. *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 West Bohemia in Pilsen](https://md.research.elamur.ai/free/en/universities/university-of-west-bohemia-in-pilsen/) — University, Pilsen, Czechia · 14 works · 38,760 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works 🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works ## Cite Cite as: Stephen Taylor. 2023. UWB at Arabic Reverse Dictionary shared task: Computing the meaning of a gloss. Proceedings of ArabicNLP 2023. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{taylor2023uwb, title = {{UWB at Arabic Reverse Dictionary shared task: Computing the meaning of a gloss}}, author = {Stephen Taylor}, booktitle = {Proceedings of ArabicNLP 2023}, year = {2023}, doi = {10.18653/v1/2023.arabicnlp-1.40}, url = {https://doi.org/10.18653/v1/2023.arabicnlp-1.40} } ``` - [← previous Using Masked Language Model Probabilities of Connectives for Stance Detection in English Discourse](https://md.research.elamur.ai/free/en/papers/stodden-2023-masked-language-model-probabilities-connectives/) - [next → When Agreement Isn't Agreement: Deflected Agreement in Najdi Arabic as Aspect](https://md.research.elamur.ai/free/en/papers/rouillier-2023-when-agreement-isn-t-agreement/) - [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/14/) ## References 14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Carl Allen and Timothy Hospedales. 2019. Analogies explained: Towards understanding word embeddings. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 223–231. PMLR. [↗](https://doi.org/10.48550/arxiv.1901.09813) 2. Waad AlShammari. 2023. ArReverseDictionary github site. [find](https://scholar.google.com/scholar?q=Waad+AlShammari%2E+2023%2E+ArReverseDictionary+github+site%2E) 3. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. [A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings](https://md.research.elamur.ai/free/en/papers/artetxe-2018-vecmap-unsupervised/). In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. 4. Tomáš Brychcín, Stephen Taylor, and Lukáš Svoboda. 2019. Cross-lingual word analogies using linear transformations between semantic spaces. Expert Systems with Applications, 135:287–295. [↗](https://doi.org/10.1016/j.eswa.2019.06.021) 5. Alex Gittens, Dimitris Achlioptas, and Michael W. Mahoney. 2017. Skip-gram - Zipf + uniform = vector additivity. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 69–76, Vancouver, Canada. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Alex+Gittens%2C+Dimitris+Achlioptas%2C+and+Michael+W%2E+Mahoney%2E+2017%2E+Skip%2Dgram+%2D+Zipf+%2B+uniform+%3D+vector+additivity%2E+In+Proceedings+of+the+55th+Annual+Meeting+of+the+Association+for+Computational+Linguist) 6. Bin Li, Yixuan Weng, Fei Xia, Shizhu He, Bin Sun, and Shutao Li. 2022. LingJing at SemEval-2022 task 1: Multi-task self-supervised pre-training for multilingual reverse dictionary. In Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 29–35, Seattle, United States. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Bin+Li%2C+Yixuan+Weng%2C+Fei+Xia%2C+Shizhu+He%2C+Bin+Sun%2C+and+Shutao+Li%2E+2022%2E+LingJing+at+SemEval%2D2022+task+1%3A+Multi%2Dtask+self%2Dsupervised+pre%2Dtraining+for+multilingual+reverse+dictionary%2E+In+Proceedings+of+t) 7. Timothee Mickus, Kees Van Deemter, Mathieu Constant, and Denis Paperno. 2022. [Semeval-2022 task 1: CODWOE – comparing dictionaries and word embeddings](https://md.research.elamur.ai/free/en/papers/mickus-2022-codwoe-dictionaries-embeddings/). In Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1–14, Seattle, United States. Association for Computational Linguistics. 8. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. [Efficient estimation of word representations in vector space](https://md.research.elamur.ai/free/en/papers/mikolov-2013-efficient-estimation-word-representations-vector/). In Proceedings of the International Conference on Learning Representations. arXiv1301.3781. [↗](https://arxiv.org/abs/1301.3781) 9. Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. 2013b. [Linguistic regularities in continuous space word representations](https://md.research.elamur.ai/free/en/papers/mikolov-2013-linguistic-regularities-continuous-space-word/). In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 746–751, Atlanta, Georgia. Association for Computational Linguistics. 10. Jeff Mitchell and Mirella Lapata. 2010. Composition in distributional models of semantics. Cognitive Science, 34:1388–1429. [↗](https://doi.org/10.1111/j.1551-6709.2010.01106.x) 11. Driss Namly. 2015. LMF contemporary arabic dictionary. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University. [find](https://scholar.google.com/scholar?q=Driss+Namly%2E+2015%2E+LMF+contemporary+arabic+dictionary%2E+LINDAT%2FCLARIAH%2DCZ+digital+library+at+the+Institute+of+Formal+and+Applied+Linguistics+%28%C3%9AFAL%29%2C+Faculty+of+Mathematics+and+Physics%2C+Charles+Universi) 12. Stephen Taylor and Tomáš Brychcín. 2018. The representation of some phrases in arabic word semantic vector spaces. Open Computer Science, 8(1):182–193. [↗](https://doi.org/10.1515/comp-2018-0017) 13. Mohamed A Zahran, Ahmed Magooda, Ashraf Y Mahgoub, Hazem Raafat, Mohsen Rashwan, and Amir Atyia. 2015a. Word representations in vector space and their applications for arabic. In International Conference on Intelligent Text Processing and Computational Linguistics, page 430–443. Springer. [find](https://scholar.google.com/scholar?q=Mohamed+A+Zahran%2C+Ahmed+Magooda%2C+Ashraf+Y+Mahgoub%2C+Hazem+Raafat%2C+Mohsen+Rashwan%2C+and+Amir+Atyia%2E+2015a%2E+Word+representations+in+vector+space+and+their+applications+for+arabic%2E+In+International+Confere) 14. Mohamed A. Zahran, Mohsen A. Rashwan, and Hazem Raafat. 2015b. Cross lingual lexical substitution using word representation in vector space. In FLAIRS. [find](https://scholar.google.com/scholar?q=Mohamed+A%2E+Zahran%2C+Mohsen+A%2E+Rashwan%2C+and+Hazem+Raafat%2E+2015b%2E+Cross+lingual+lexical+substitution+using+word+representation+in+vector+space%2E+In+FLAIRS%2E) *Updated 1 November 2026*