Pith. sign in

REVIEW 5 major objections 6 minor 24 references

Advancing Arabic Reverse Dictionary Systems: A Transformer-Based Approach with Dataset Construction Guidelines

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims ARBERTv2, paired with a shrinking-layer semi-encoder network, attains the best Arabic reverse-dictionary rank (0.0644).

desk verdict Headline rank repeats their own 2024 result, the rank equations are incoherent as written, and the dataset-quality guidelines are the only clearly new value. read the letter →

arxiv 2504.21475 v1 pith:EN2OFOV6 submitted 2025-04-30 cs.CL cs.AIcs.LG

Serry Sibaee , Samar Ahmed , Abdullah Al Harbi , Omer Nacar , Adel Ammar , Yasser Habashi , Wadii Boulila This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords ArabicreversedictionarytransformermodelsARBERTv2semanticsearchsemi-encodernetworklexicographicdefinitionsdatasetconstructionstandardsembedding-basedretrieval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a simple transformer-plus-regression pipeline can retrieve Arabic words from plain definitions, and that the choice of Arabic-specific definition encoder matters more than architectural complexity. The authors claim that ARBERTv2, paired with a semi-encoder network whose hidden layers shrink geometrically, achieves the best ranking score of 0.0644 among the tested models on the Arabic reverse-dictionary task. They also argue that the training data itself is the main bottleneck and propose eight standards for writing definitions so future datasets support better retrieval. If the claim holds, Arabic speakers gain a practical tool for tip-of-the-tongue searches and dataset builders get a concrete checklist.

What carries the argument

The load-bearing object is the learned map $m: \mathbb{R}^d \to \mathbb{R}^b$ from definition-embedding space to word-embedding space, realized as a semi-encoder: a four-layer feedforward network with hidden widths $8s, 4s, 2s, s$, where $s=d=256$, GELU activations, dropout between 0.2 and 0.4, and the AdamW optimizer at learning rate $10^{-4}$. The network is trained with mean squared error loss on pairs $(f(\text{def}_i), e(\text{word}_i))$, and a definition is judged correct when its mapped vector has top cosine similarity to the true word embedding among all candidates. This geometric alignment is what carries the retrieval; no external lexicon or morphological rules are used at inference time.

What would settle it

Run the released RDTL pipeline on the official test split and evaluate with the official rank metric; if the average rank of the true word differs materially from 0.0644, or if applying Equations (5)-(6) literally assigns the same rank to every candidate, the headline result does not hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that an Arabic-specific pretrained encoder is the decisive ingredient in Arabic reverse-dictionary retrieval, and that ARBERTv2 preserves the relative semantic distances needed for ranking better than any tested alternative. With a semi-encoder network whose hidden layers decrease geometrically from $8s$ to $s$ ($s=256$), trained against 256-dimensional AraELECTRA word embeddings using mean squared error loss, ARBERTv2 achieves the best ranking score of $0.0644$ on the shared-task test set, ahead of the next-best OpenAI large embedding at $0.077$ and above the baseline score of $0.84$. The paper also claims that inspecting more than 200 randomly selected definitions exposes recurring quality problems, and it condenses those findings into eight standards for constructing Arabic reverse-dictionary datasets.

Load-bearing premise

The load-bearing premise is that the reported ranking score was computed with the official shared-task rank procedure, not literally with the paper's own rank equations, since those equations compare each score to itself and would not discriminate between candidates.

Editorial extensions

If this is right

  • With ARBERTv2 as the definition encoder, a new Arabic reverse-dictionary deployment can start from the released RDTL pipeline and expect the best tested rank of 0.0644 rather than searching across multilingual embeddings.
  • If the eight definition-writing standards are adopted, future dataset builders can avoid morphology-only entries, ambiguous pronouns, and field-specific definitions that the analysis found in more than 30 percent of sampled cases.
  • The formal abstraction of reverse dictionary as a learned map between definition and word embedding spaces gives future system builders a single training and evaluation target.
  • The modular library lets researchers swap encoders and retrain, so the same architecture can be tested on new Arabic datasets without re-implementing the pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension the paper does not run is to rebuild a small gold-standard Arabic definition set under its eight standards and retrain the same architecture on it; fixed-capacity improvement over the raw dataset would isolate how much of the ranking gains come from data quality rather than encoder choice.
  • Because the printed rank equations are internally inconsistent, an editor's check would be to re-implement the official evaluator and recompute Table 3; that would confirm whether the 0.0644 figure and the baseline comparisons are directly comparable.
  • The superiority of Arabic-specific encoders is demonstrated on Arabic definitions and Arabic word embeddings; whether it persists under cross-lingual or dialectal queries is not tested.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This paper presents an Arabic reverse dictionary (RD) system built around a proposed "semi-encoder" feed-forward network with geometrically decreasing hidden layers, trained to map definition embeddings to target word embeddings. The authors evaluate several pre-trained transformer encoders, report ARBERTv2 as the best model with a ranking score of 0.0644, and claim state-of-the-art performance. The paper also proposes eight standards for Arabic lexicographic definition quality, presents a qualitative dataset analysis, and releases a Python library called RDTL.

Significance. If the reported result were substantiated, the paper would offer a useful benchmark for Arabic reverse dictionary research and a practical open-source library. The breadth of compared encoders and the qualitative dataset-quality analysis are valuable, and releasing RDTL is a constructive step toward reproducibility. However, the central quantitative claim is not verifiable as written: the ranking metric is defined incoherently, the source of the 84,000 augmented samples is unspecified, no error bars or significance tests are provided, and the headline number 0.0644 coincides exactly with a value the authors reported in their own prior KSAA-CAD publication. The claimed state-of-the-art result therefore cannot be accepted on the evidence in this manuscript.

major comments (5)
  1. [Section 3.2.1, Eqs. (5)-(7)] The definition of the ranking metric is incoherent. M_w is defined as the set of metric scores for a single word w, and Eqs. (5)-(6) sum indicator functions over that same set; this does not compare a candidate's score with the scores of other candidates. A retrieval rank requires a comparison between S_w and the scores of all candidates in a defined vocabulary V, e.g., R_w = sum_{v in V} 1(S_v >= S_w). The vocabulary V is never specified beyond its appearance in Eq. (4), and the official KSAA-CAD evaluation protocol is not cited. Consequently, the rank column of Table 3 and the headline 0.0644 cannot be interpreted or reproduced from the text.
  2. [Section 3.1] The approximately 84,000 additional training samples are said to come from an "external source" with no citation, description, or construction details. This is load-bearing because the augmented set is larger than the original 31,372 training samples and could dominate training; without knowing its provenance, domain, and overlap with the test set, the experimental results are not reproducible.
  3. [Section 5.1 and Table 3] All results are reported as single runs with no standard deviations, confidence intervals, or significance tests. The text repeatedly uses terms such as "significant performance gap" and "significantly impact" without statistical support. Given the small rank differences among several models (e.g., ARBERTv2 at 0.0644 versus OpenAI large at 0.077), the claim that ARBERTv2 is the best model is not established.
  4. [Section 2.3 and abstract] The reported best rank of 0.0644 is identical to the rank the authors report for their previous system in Sibaee et al. (2024), described in Section 2.3 as achieving "the best rank and an MSE score of 0.0644 and 0.059." The manuscript provides no explanation for this coincidence and no independent evidence that the current ARBERTv2 configuration was evaluated with a correct ranking protocol. Because the abstract's state-of-the-art claim depends directly on this number, the relationship between the current result and the prior publication must be clarified with an independen computation or an explicit statement of what is new.
  5. [Section 3.2 and Figure 1] The manuscript does not specify how each transformer encoder is applied to definitions, including pooling strategy, maximum sequence length, and whether the encoder is fine-tuned or frozen. The claim that the architecture is consistent across all experiments therefore cannot be checked, and the released library is not documented with the evaluation script used to compute the rank column. These details are necessary for reproducibility of the central experiments.
minor comments (6)
  1. [Section 3.2] There are typographical errors in "this architicture is chosen after empirical expirements from Sibaee et al. (2023)"; the sentence should also clarify why the authors' own prior experience is the only basis for the architecture choice.
  2. [Section 5.2 and Figure 2] The annotation procedure for the 200+ randomly selected samples is not described: the number of annotators, inter-annotator agreement, and sample selection method are missing. Figure 2's axes are labeled "Class" and "Count of accuracy," which does not make clear whether the scores are 1-5 quality ratings or accuracy classes.
  3. [Section 4.1 and Table 3] The model "baselineM_elct" is never defined in the text; the baseline should be clearly described in the methodology so readers know what it represents.
  4. [Section 3.2.1, Eq. (7)] The notation R_i,w is used before it is defined, and the phrase "aggregated across users" suggests a user study that is not presented anywhere in the paper.
  5. [Section 2.3] The sentence describing Sibaee et al. (2024) is grammatically ambiguous: "the best rank and an MSE score of 0.0644 and 0.059" should be rephrased as "a rank of 0.0644 and an MSE of 0.059."
  6. [Title and abstract] The paper is described as proposing a "Transformer-Based Approach," but the actual novelty is a fully connected semi-encoder network on top of existing transformer embeddings; the contribution boundary should be stated more precisely in the title and abstract.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline rank score is definitionally self-referential and traces to the same team's prior semi-encoder result; central SOTA claim partly reduces to self-citation.

  1. self definitional [Section 3.2.1, Equations (5)-(7)]
    "For each word w, a set of metric scores M_w = {M_1(w), M_2(w), …, M_k(w)} is computed, where each M_i(w) represents a different evaluation metric (e.g., MSE, cosine similarity, rank). The rank transformation is defined as: R_w = ∑_{x∈M_w} 1(x ≥ M_w) if the metric is to be maximized, (5) R_w = ∑_{x∈M_w} 1(x ≤ M_w) if the metric is to be minimized. (6)"

    The rank score R_w is defined as a sum over the set M_w, and M_w is specified to contain 'rank' itself as one of its elements. There is no independent definition of the rank element or of the candidate vocabulary V used in Eq. (4), so Equations (5)-(7) cannot define a retrieval rank: the quantity being computed is an input to its own computation. The headline value 0.0644 therefore cannot be verified from the paper's own formal specification; it either comes from an unstated official protocol or from following a self-referential definition.

  2. self citation load bearing [Section 2.3; Section 3.2; Table 3; Section 5.1]
    "Sibaee et al. (2024) used a semi-encoder structure for reverse dictionary tasks, ... Using ELECTRA with this model, their system achieved the best rank and an MSE score of 0.0644 and 0.059. ... this architicture is chosen after empirical expirements from Sibaee et al. (2023). ... ARBERTv2 achieves the best rank metric (0.0644), which is the primary evaluation criterion for our study."

    The load-bearing pieces of the central SOTA claim are supplied by the same authors' earlier papers. The semi-encoder architecture is justified solely as 'chosen after empirical experiments from Sibaee et al. (2023)', and the headlined rank 0.0644 is numerically identical to the rank previously reported for the same team's semi-encoder with AraBERTv2 in Sibaee et al. (2024). The paper presents this as a new state-of-the-art finding rather than deriving it from new external evidence, so the central claim reduces to a self-citation chain plus the new, unverifiable rank equations.

full rationale

The paper is not entirely circular: the dataset-quality standards (Section 5.2) and the RDTL library are independent practical contributions, and Table 3 does contain a fresh comparison of multiple embedding models under one fixed architecture. However, the paper's central claim—the state-of-the-art rank of 0.0644—is compromised. First, Section 3.2.1's formal definition of rank is self-referential: M_w is said to include 'rank' as one of its metric scores, and Equations (5)-(7) define R_w as a sum over that same set, with no definition of the candidate set V or the rank element. No retrieval rank can be recovered from these equations, so the headline number is not derivable from the paper's own methodology. Second, the architecture and the exact score 0.0644 trace to the same team's prior papers (Sibaee et al., 2023, 2024), which are cited as the justification for the architecture and which already reported the same numeric rank. The new experiments compare encoders but do not establish SOTA against an externally specified protocol; the 'state-of-the-art' label rests on the prior self-cited result. This is partial circularity rather than total: the dataset analysis and library are independent, and the encoder comparison could in principle be reproduced if the official KSAA-CAD evaluation were supplied. Score 6 reflects that the central metric reduces by construction/self-citation, while non-central contributions retain independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests mainly on inherited hyperparameters and unexamined domain assumptions about embedding spaces and data quality. No genuinely new formal entities are postulated; the only novel contributions are qualitative guidelines derived from a small manual review.

free parameters (4)
  • Hidden layer sizes (geometric decay) = 4 layers: 2048, 1024, 512, 256 (8s, 4s, 2s, s with s=256)
    Chosen based on prior empirical experiments from Sibaee et al. (2023); no independent justification in this paper.
  • Dropout rate = d in [0.2, 0.4], exact value not reported
    Adjustable dropout to mitigate overfitting; the actual value per experiment is not specified.
  • Learning rate = 1e-4
    Set because higher values led to instability in previous experiments; not tuned in this paper.
  • Augmentation sample count = 84,000 samples
    Number of additional samples from an unnamed external source; the choice is not justified or described.
assumptions (4)
  • domain assumption Cosine similarity in the pretrained embedding space is a valid proxy for word meaning retrieval.
    The ranking evaluation assumes that higher cosine similarity between predicted and candidate word embeddings corresponds to better reverse dictionary performance.
  • domain assumption The fixed ELECTRA word embeddings (e: Word -> R^256) provide an adequate target space for word meanings.
    The network is trained to predict these embeddings; the quality or limitations of this target space are not analyzed.
  • domain assumption The 84k externally sourced augmented samples are compatible with the KSAA-RD dataset distribution and do not leak into the test set.
    No description, source, or overlap analysis is provided for the augmented data.
  • domain assumption Mean squared error on embeddings is an appropriate training objective for the mapping from definitions to words.
    Standard regression loss used without discussion; the paper does not evaluate alternative losses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Arabic Reverse Dictionary Systems: A Transformer-Based Approach with Dataset Construction Guidelines." pith.science (2026). https://pith.science/paper/EN2OFOV6

@misc{pith2026250421475,
  author       = {Pith},
  title        = {Pith review of: Advancing Arabic Reverse Dictionary Systems: A Transformer-Based Approach with Dataset Construction Guidelines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EN2OFOV6}},
  note         = {Machine review of arXiv:2504.21475}
}
read the original abstract

This study addresses the critical gap in Arabic natural language processing by developing an effective Arabic Reverse Dictionary (RD) system that enables users to find words based on their descriptions or meanings. We present a novel transformer-based approach with a semi-encoder neural network architecture featuring geometrically decreasing layers that achieves state-of-the-art results for Arabic RD tasks. Our methodology incorporates a comprehensive dataset construction process and establishes formal quality standards for Arabic lexicographic definitions. Experiments with various pre-trained models demonstrate that Arabic-specific models significantly outperform general multilingual embeddings, with ARBERTv2 achieving the best ranking score (0.0644). Additionally, we provide a formal abstraction of the reverse dictionary task that enhances theoretical understanding and develop a modular, extensible Python library (RDTL) with configurable training pipelines. Our analysis of dataset quality reveals important insights for improving Arabic definition construction, leading to eight specific standards for building high-quality reverse dictionary resources. This work contributes significantly to Arabic computational linguistics and provides valuable tools for language learning, academic writing, and professional communication in Arabic.

Figures

Figures reproduced from arXiv: 2504.21475 by the authors.

Figure 1
Figure 1. Overview of the encoder-decoder flow with embedding and target loss. 3.2.1. Inference and Ranking-Based Evaluation Given a test definition def′ , we retrieve the most relevant word by ranking candidate words based on a similarity metric. • Similarity Computation We define a similarity function 𝑆 ∶ ℝ𝑏 × ℝ𝑏 → ℝ, such as cosine similarity: 𝑆(𝑥, 𝑦) = 𝑥 ⋅ 𝑦 ‖𝑥‖‖𝑦‖ (3) For each candidate word 𝑤 ∈ 𝑉 , we compute: 𝑆𝑤 = 𝑆(𝑚(… view at source ↗
Figure 2
Figure 2. Accuracy distribution cases include ¨ñ JK [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Summary of RD writing standards. unambiguous references, general before specific meanings, domain specification, avoidance of figurative language, elimination of redundancy, comprehensive descriptions over mere synonyms, and logical structure. The modular RDTL library we developed provides researchers and developers with configurable training pipelines for RD tasks, facilitating further innovation in this field. Fut… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 14 canonical work pages

  1. [1]

    Abdul-Mageed,M.,Elmadany,A.,Nagoudi,E.M.B.,2021. ARBERT&MARBERT:DeepbidirectionaltransformersforArabic,in:Zong,C.,Xia, F.,Li,W.,Navigli,R.(Eds.),Proceedingsofthe59thAnnualMeetingoftheAssociationforComputationalLinguisticsandthe11thInternational JointConferenceonNaturalLanguageProcessing(Volume1:LongPapers),AssociationforComputationalLinguistics,Online.pp....

    doi:10.18653/v1/2021.acl-long.551 2021

  2. [4]

    Baleegh at KSAA-CAD 2024: Towards enhancing Arabic reverse dictionaries, in: Habash, N., Bouamor, H., Eskander, R., Tomeh, N., Abu Farha, I., Abdelali, A., Touileb, S., Hamed, I., Onaizan, Y., Alhafni, B., Antoun, W., Khalifa, S., Haddad, H., Zitouni, I., AlKhamissi, B., Almatham, R., Mrini, K. (Eds.), Proceedings of the Second Arabic Natural Language Pro...

    work page 2024

  3. [7]

    A unified model for reverse dictionary and definition modelling, in: He, Y., Ji, H., Li, S., Liu, Y., Chang, C.H. (Eds.), Proceedingsofthe2ndConferenceoftheAsia-PacificChapteroftheAssociationforComputationalLinguisticsandthe12thInternationalJoint ConferenceonNaturalLanguageProcessing(Volume2:ShortPapers),AssociationforComputationalLinguistics,Onlineonly.p...

    work page doi:10.18653/v1/2022.aacl-short.2 2022

  4. [8]

    Computational Intelligence and Neuroscience 2022, 6561622

    A natural language processing (nlp) evaluation on covid-19 rumour dataset using deep learning techniques. Computational Intelligence and Neuroscience 2022, 6561622. Hedderich, M.A., Yates, A., Klakow, D., de Melo, G.,

    work page 2022

  5. [10]

    PREDICT: Persian Reverse Dictionary

    Predict: Persian reverse dictionary. URL: https://arxiv.org/abs/2105.00309, arXiv:2105.00309. Mane, S.B., Patil, H., Madaswar, K., Sadavarte, P.,

    work page Pith review arXiv

  6. [11]

    URL:https://arxiv.org/ abs/2204.10181, arXiv:2204.10181

    Wordalchemy: A transformer-based reverse dictionary. URL:https://arxiv.org/ abs/2204.10181, arXiv:2204.10181. Méndez, O., Calvo, H., Moreno-Armendáriz, M.A.,

    work page arXiv

  7. [13]

    (Eds.), Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), Association for Computational Linguistics, Seattle, United States

    Semeval-2022 task 1: CODWOE – comparing dictionaries and word embeddings, in: Emerson, G., Schluter, N., Stanovsky, G., Kumar, R., Palmer, A., Schneider, N., Singh, S., Ratan, S. (Eds.), Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), Association for Computational Linguistics, Seattle, United States. pp. 1–14. URL: ht...

    doi:10.18653/v1/2022.semeval-1.1 2022

  8. [14]

    URL: https: //arxiv.org/abs/2407.21139, arXiv:2407.21139

    Enhancing semantic similarity understanding in arabic nlp with nested embedding learning. URL: https: //arxiv.org/abs/2407.21139, arXiv:2407.21139. Nacar, O., Koubaa, A., Sibaee, S.T., Ghouti, L.,

    arXiv

Show all 24 references
  1. [16]

    URL: https://aclanthology.org/2023.arabicnlp-1.42/, doi:10.18653/v1/2023.arabicnlp-1.42

    Abed at KSAA-RD shared task: Enhancing Arabic word embedding with modified BERT multilingual, in: Sawaf, H., El- Beltagy,S.,Zaghouani,W.,Magdy,W.,Abdelali,A.,Tomeh,N.,AbuFarha,I.,Habash,N.,Khalifa,S.,Keleg,A.,Haddad,H.,Zitouni,I.,Mrini, K.,Almatham,R.(Eds.),ProceedingsofArabic...

    2023 doi

  2. [18]

    URL:https://arxiv

    Making monolingual sentence embeddings multilingual using knowledge distillation. URL:https://arxiv. org/abs/2004.09813, arXiv:2004.09813. Roman, M., Shahid, A., Khan, S., Koubaa, A., Yu, L.,

    2004 arXiv

  3. [19]

    Ieee Access 9, 9982–9995

    Citation intent classification using word embedding. Ieee Access 9, 9982–9995. Sibaee,S.,Ahmad,S.,Khurfan,I.,Sabeeh,V.,Bahaaulddin,A.,Belhaj,H.,Alharbi,A.,2023.QamosyatArabicreversedictionarysharedtask:Semi decoderarchitectureforreversedictionarywithSBERTencoder,in:Sawaf,H.,El...

    2023

  4. [20]

    (Eds.), Proceedings of the Second Arabic Natural Language Processing Conference, Association for Computational Linguistics, Bangkok, Thailand

    ASOS at KSAA-CAD 2024: One embedding is all you need for your dictionary, in: Habash, N., Bouamor, H., Eskander, R., Tomeh, N., Abu Farha, I., Abdelali, A., Touileb, S., Hamed, I., Onaizan, Y., Alhafni, B., Antoun, W., Khalifa, S., Haddad, H., Zitouni, I., AlKhamissi, B., Alma...

    2024

  5. [21]

    (Eds.), Proceedings of ArabicNLP 2023, Association for Computational Linguistics, Singapore (Hybrid)

    UWB at Arabic reverse dictionary shared task: Computing the meaning of a gloss, in: Sawaf, H., El-Beltagy, S., Zaghouani, W., Magdy, W., Abdelali, A., Tomeh, N., Abu Farha, I., Habash, N., Khalifa, S., Keleg, A., Haddad, H., Zitouni, I., Mrini, K., Almatham, R. (Eds.), Proceed...

    2023 doi

  6. [23]

    (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, Association for Computational Linguistics, Online

    BERT for monolingual and cross-lingual reverse dictionary, in: Cohn, T., He, Y., Liu, Y. (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, Association for Computational Linguistics, Online. pp. 4329–4338. URL: https://aclanthology.org/2020.findings...

    2020 arXiv

  7. [24]

    URL:https://arxiv.org/abs/1912

    Multi-channel reverse dictionary model. URL:https://arxiv.org/abs/1912. 08441, arXiv:1912.08441. Sibaee et al.: Preprint submitted to Elsevier Page 12 of 12

    1912 arXiv

  8. [2013]

    A reverse dictionary based on semantic analysis using wordnet, in: Advances in ArtificialIntelligenceandItsApplications:12thMexicanInternationalConferenceonArtificialIntelligence,MICAI2013,MexicoCity,Mexico, November 24-30, 2013, Proceedings, Part I 12, Springer. pp. 275–285. ...

    2013

  9. [2016]

    (Eds.), Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, The COLING 2016 Organizing Committee, Osaka, Japan

    Implementing a reverse dictionary, based on word definitions, using a node-graph architecture, in: Matsumoto, Y., Prasad, R. (Eds.), Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, The COLING 2016 Organizing Committ...

    2016

  10. [2019]

    URL:https: //arxiv.org/abs/1904.01451, arXiv:1904.01451

    Using multi-sense vector embeddings for reverse dictionaries. URL:https: //arxiv.org/abs/1904.01451, arXiv:1904.01451. Malekzadeh, A., Gheibi, A., Mohades, A.,

    1904 arXiv

  11. [2020]

    (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for ComputationalLinguistics,Online.pp.175–181

    WantWords: An open-source online reverse dictionary system, in: Liu, Q., Schlangen, D. (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for ComputationalLinguistics,Online.pp.175–181. URL: https:...

    2020 doi

  12. [2021]

    (Eds.), Proceedings of the Sixth Arabic Natural Language Processing Workshop, Association for Computational Linguistics, Kyiv, Ukraine (Virtual)

    AraELECTRA: Pre-training text discriminators for Arabic language understanding, in: Habash, N., Bouamor, H., Hajj, H., Magdy, W., Zaghouani, W., Bougares, F., Tomeh, N., Abu Farha, I., Touileb, S. (Eds.), Proceedings of the Sixth Arabic Natural Language Processing Workshop, As...

    2021

  13. [2022]

    (Eds.), Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), Association for Computational Linguistics, Seattle, United States

    BL.Research at SemEval-2022 task 1: Deep networks for reverse dictionary using embeddings and LSTM autoencoders, in: Emerson, G., Schluter, N., Stanovsky, G., Kumar, R., Palmer, A., Schneider, N., Singh, S., Ratan, S. (Eds.), Proceedings of the 16th International Workshop on S...

    2022 doi

  14. [2023]

    KSAA-RD shared task: Arabic reverse dictionary, in: Sawaf, H., El-Beltagy, S., Zaghouani, W., Magdy, W., Abdelali, A., Tomeh, N., Abu Farha, I., Habash, N., Khalifa,S.,Keleg,A.,Haddad,H.,Zitouni,I.,Mrini,K.,Almatham,R.(Eds.),ProceedingsofArabicNLP2023,AssociationforComputation...

    2023 doi

  15. [2024]

    (Eds.), Proceedings of the Second Arabic Natural Language Processing Conference, Association for Computational Linguistics, Bangkok, Thailand

    MISSION at KSAA-CAD 2024: AraT5 with Arabic reverse dictionary, in: Habash, N., Bouamor, H., Eskander, R., Tomeh, N., Abu Farha, I., Abdelali, A., Touileb, S., Hamed, I., Onaizan, Y., Alhafni, B., Antoun, W., Khalifa, S., Haddad, H., Zitouni, I., AlKhamissi, B., Almatham, R., ...

    2024

  16. [2025]

    URL:https://arxiv.org/abs/2502.07972, arXiv:2502.07972

    Training sparse mixture of experts text embedding models. URL:https://arxiv.org/abs/2502.07972, arXiv:2502.07972. Pilehvar,M.T.,2019. Ontheimportanceofdistinguishingwordmeaningrepresentations:Acasestudyonreversedictionarymapping,in:Burstein, J., Doran, C., Solorio, T. (Eds.), ...

    2019 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.