PaperlayerView exact revision on arXiv

Source-linked AI summary

Opportunities and Challenges of Natural Language Processing for Low-Resource Senegalese Languages in Social Science Research

Derguene Mbaye, Tatiana D. P. Mbengue, Madoune R. Seye, Moussa Diallo, Mamadou L. Ndiaye, Dimitri S. Adjanohoun, Cheikh S. Wade, Djiby Sow, Jean-Claude B. Munyaka, Jerome Chenal

arXiv:2601.09716v2cs.CL

TL;DR

African languages remain underrepresented in NLP, and Senegal’s six national languages lack an exclusive comprehensive survey. This paper synthesizes their NLP resources and challenges, and finds broad coverage for Wolof but only keyword-spotting coverage shared by the other five languages.

  • Problem

    African languages remain underrepresented in NLP, and the paper identifies a lack of comprehensive coverage of Senegal’s six national languages.

    Source: §1 · p. 2

  • Method

    The paper surveys NLP progress and challenges across the six languages and task spectrum, and compiles publicly accessible resources in a centralized GitHub repository.

    Source: §1 · p. 2, §8 · p. 28

  • Results

    Wolof has some coverage across all task categories, while the other five languages collectively have full coverage only in keyword spotting.

    Source: Cross-Language and Cross-Task Coverage… · pp. 23–24

  • Takeaways & Limitations

    The paper identifies NLP’s potential to accelerate ethnographic analysis and make social science research outputs more inclusive and timely.

    Source: §7.2 · p. 27

    Wolof NER approaches have not been directly benchmarked on the same test split, and the small MasakhaNER test set makes reported F1 scores variable.

    Source: §5.3 · p. 10

Abstract

from arXiv · show

Natural Language Processing (NLP) is rapidly transforming research methodologies across disciplines, yet African languages remain largely underrepresented in this technological shift. This paper provides the first comprehensive overview of NLP progress and challenges for the six national languages officially recognized by the Senegalese Constitution: Wolof, Pulaar, S\'er\`ere, Diola, Mandingue, and Sonink\'e. We synthesize linguistic, socio-technical, and infrastructural factors that shape their digital readiness and identify gaps in data, tools, and benchmarks. Building on existing initiatives and research works, we analyze ongoing efforts in various tasks, covering both text and speech modalities. We also provide a centralized GitHub repository that compiles publicly accessible resources for a range of NLP tasks across these languages, designed to facilitate collaboration and reproducibility. A special focus is devoted to the application of NLP to the social sciences, where multilingual transcription, translation, and retrieval pipelines can significantly enhance the efficiency and inclusiveness of field research. The paper concludes by outlining a roadmap toward sustainable, community-centered NLP ecosystems for Senegalese languages, emphasizing ethical data governance, open resources, and interdisciplinary collaboration.

1. Introduction

NLP advances have largely bypassed African languages, while Senegal’s six constitutionally recognized national languages remain underrepresented in digital and scientific NLP. The paper surveys their resources and challenges, connecting this landscape to social science research and a publicly accessible resource repository.

1. Introduction

  • African languages face a pronounced NLP performance gap: Pulaar, Wolof, and other widely spoken languages underperform English across reasoning, generation, and classification tasks.The paper describes this as a newer digital divide concerning languages’ representation and processing by AI systems.Source: §1 · p. 2
  • Senegalese languages’ exclusion from NLP risks technological marginalization and leaves opportunities for locally grounded social science research untapped.Social science fieldwork often involves multilingual interviews, focus groups, and recordings that are resource-intensive to transcribe, translate, and analyze; applying NLP requires robust resources, interoperable tools, and sustainable community infrastructure.Source: §1 · p. 2
  • The survey covers all six national languages, maps progress across text and speech tasks, and frames findings around social science field research.It also introduces a publicly accessible GitHub repository of datasets, benchmarks, and tools, intended to grow through community contributions.Source: §1 · p. 2
  • Earlier surveys cover African-language research broadly or individual country ecosystems, but none focuses exclusively on Senegal’s national languages.The paper distinguishes its scope from language inventories that lack its task-by-task analysis and social science framing.Source: §1 · p. 2

2. Survey Methodology

The survey assembled relevant work through a three-phase search designed to cover both conventional publication venues and community channels. It filtered the resulting pool using explicit inclusion and exclusion criteria.

2. Survey Methodology

  • Keyword queries paired the six language names and aliases with NLP task terms; forward and backward citation tracking surfaced additional relevant work.Citation tracking helped identify regional-venue and workshop papers not prominent in keyword searches.Source: §2 · p. 2
  • The search combined keyword-driven Google Scholar queries, citation tracking, and author-centric searches with grey-literature targeting.The process then deduplicated and screened the raw pool against explicit criteria.Source: Figure 2 · p. 3
  • Author publication lists and community sources, including blogs, reports, and project pages, added work beyond traditional bibliographies.Included work produced or evaluated a resource, tool, model, or benchmark for at least one national language; French- or Arabic-only pivot work and inaccessible internal reports were excluded.Source: §2 · p. 3

3. Sociolinguistic and Linguistic Features

Senegal’s six national languages span varied linguistic, orthographic, and sociolinguistic contexts that shape NLP readiness. Their official recognition and standardized orthographies coexist with French-dominant education, uneven materials, spelling variation, and dialectal diversity; language overview details also point to localized distributions and distinct language families.

3. Sociolinguistic and Linguistic Features

  • Although the national languages have standardized orthographies, French remains the sole language of education and administration, while newer bilingual education efforts face resource and coordination challenges.MOHEBS is expanding across 13 of 16 educational regions, but inconsistent spelling and substantial dialectal variation remain computational challenges.Source: §3 · pp. 3–4

3.1. Language Digraphs

  • Ajami and Latin writing practices largely serve different contexts; AjamiXTranslit offers Latin–Ajami resources and models, though its largest Senegalese-language corpus has fewer than 70 rows.The cited corpus covers Wolof, Pulaar, and Soninké in Ajami script.Source: §3.1 · pp. 4–5

3.2.1. Wolof

  • Wolof’s broad reach and early digitalization make it Senegal’s most resourced NLP language, but orthographic variation, morphophonemic alternations, and urban code-switching complicate processing.Senegalese and Gambian Wolof are treated as distinct languages, and urban Wolof includes French borrowing.Source: §3.2.1 · p. 5

3.2.2. Pulaar / Fula

  • Pulaar’s complex noun classes, verbal extensions, and consonant mutation complicate processing; its transnational presence offers scope for regional work, while cross-border orthographic harmonization remains incomplete.Senegal’s dominant variety is associated with Fouta Toro, alongside other dialects spoken in the country.Source: §3.2.2 · p. 5

3.2.3. Sérère (Sereer)

  • Sérère has an official writing system based on Seereer Siin, yet remains severely underrepresented in digital corpora despite a substantial speaker base.The term Sérère can also refer to culturally similar populations in eastern Senegal who speak Cangin languages.Source: §3.2.3 · p. 5

3.2.4. Diola (Joola)

  • Joola’s morphosyntactic diversity, dialect fragmentation, and limited orthographic standardization challenge corpus alignment; its use is primarily oral, including on local radio.Joola is mainly spoken in Casamance, and documentation projects are beginning to address the gap.Source: §3.2.4 · pp. 5–6

3.2.5. Mandingue (Mandinka)

  • Mandingue’s stable morphology and regional presence may support transfer learning from related Mande languages such as Bambara.Its described features include high–low tone, little morphological inflection, no grammatical gender, and rigid word order.Source: §3.2.5 · p. 6

3.2.6. Soninké

  • Soninké has limited digital resources and no dominant standardized variety; its cross-border use and structured morphology are presented as promising for multilingual modeling.It is tonal and has little mutual intelligibility with other Mande languages, while its dialect varieties are described as mutually intelligible.Source: §3.2.6 · p. 6

4. Low-Resource Context and Data Availability

Senegal’s NLP ecosystem is emerging through local, regional, and international activity, alongside institutional and community initiatives. Publicly known resources remain unevenly distributed across languages, and the survey provides a separate, expanding GitHub inventory.

4. Low-Resource Context and Data Availability

  • The ecosystem developed from a local-only period in 2010–2019 to regional collaboration around 2021 and multilateral acceleration from 2022.The later acceleration followed Wolof and Pulaar’s inclusion in large-scale multilingual benchmarks.Source: §4 · p. 6
  • National AI strategy, academic institutions, GalsenAI, and Masakhane contribute to resource development, training, and collaborative multilingual modeling.Survey publications are split among local (42%), regional (19%), and international (39%) origins; grey-literature contributions are disproportionately local.Source: §4 · p. 6
  • Publicly known datasets, corpora, and tools are imbalanced between Wolof and other national languages; a separate GitHub document tracks additional sources.The inventory is intended to facilitate continued tracking of datasets produced for these languages.Source: §4 · p. 6

5. Current NLP Efforts and Tasks

Local researchers lead language-specific tasks, while international teams lead infrastructure tasks; language identification remains a bottleneck, with no reliable identifier for Senegalese national languages beyond Wolof.

5. Current NLP Efforts and Tasks

  • Local researchers dominate language-specific tasks, while international teams lead ASR, MT, QA, and LLM infrastructure tasks; several categories have no local papers.Language Identification, NER, Fact Checking, and Pre-training are identified as direct priority targets.Source: §5 · p. 6, Figure 7 · p. 8
  • Funding, infrastructure access, and highly qualified talent remain constraints on local research efforts.Source: §5 · p. 7

5.1. Language Identification

  • A reliable identifier for Senegalese national languages beyond Wolof remains absent, making language identification a bottleneck for downstream tasks.The reviewed literature traces a shift from feature-based systems to massively multilingual neural models.Source: §5.1 · p. 8

5.2. Parsing & Tokenization

Wolof parsing work developed from formal linguistic tools toward integration with neural NLP, while tokenization and annotation choices continue to shape downstream constraints.

5.2. Parsing & Tokenization

  • Parsing generally assumes tokenized input, and traditional dependency or constituency parsing relies on word-level units to assign syntactic roles.Source: §5.2 · p. 8
  • The publicly available resource catalogue spans datasets and tools for Senegalese languages across tasks including machine translation, token classification, speech recognition, and language identification.Its excerpt links entries to their source repositories or download pages.Source: Table 1 · p. 9
  • Wolof parsing systems use finite-state transducers for tokenization and morphological analysis, with LFG grammars and parsing strategies handling free text and grammatical constructions.Integrating FST components addressed issues with multiword expressions, clitics, and normalization.Source: §5.2 · p. 9
  • Wolof parsing followed a rule-based-to-neural trajectory, with finite-state and LFG grammars preceding a UD treebank that enabled integration into tools such as Stanza.The UD resource also supports token-level annotation, while subword segmentation choices constrain downstream models.Source: §5.2 · pp. 9–10

5.3. Token Classification

Published NER and POS resources among Senegalese languages are limited to Wolof, while other national languages remain uncovered at the token-classification level.

5.3. Token Classification

  • Across 27 languages, including 11 absent from M-BERT such as Wolof, E-MBERT achieved an average F1 increase of 23% on new languages and 6% on languages already in M-BERT.Source: §5.3 · p. 10
  • Wolof is the only Senegalese language with published NER and POS resources; Pulaar, Sérère, Diola, Mandingue, and Soninké remain uncovered at this level.Source: §5.3 · p. 10
  • MasakhaNER transfer and E-MBERT provide complementary strategies, but have not been benchmarked on the same Wolof test split, and the small test set adds variance to reported F1 scores.A common held-out benchmark is needed to compare them and distinguish architecture, vocabulary coverage, and annotation-quality effects.Source: §5.3 · p. 10

5.4. Text Classification

Across text-classification tasks, work shifts from lexicon-based and classical methods on Wolof-French code-switched data toward multilingual-model transfer; datasets remain small and Wolof-centric.

5.4.1. Sentiment Analysis

  • French-Wolof sentiment work progressed from lexicons and word-level trigrams to dictionary-based lemmatization, context-dependent sentiment modeling, and multilingual XLM-RoBERTa.Studies lack shared held-out benchmarks with confidence intervals, and most datasets draw from a single social-media platform, limiting generalization to other registers.Source: §5.4.1 · p. 11

5.4.2. Hate Speech Detection

  • Wolof-French abusive-message work includes an annotated 2,022-tweet dataset and AbuseBERTWoFr, trained on nearly 145k tweets.The small evaluation corpus and lack of cross-dataset validation leave performance vulnerable to temporal and topical shifts across data streams and platforms.Source: §5.4.2 · p. 11

5.4.3. Intent Classification

  • Intent-classification resources include Injongo for 16 African languages and roughly 10k translated French-Wolof customer-service queries evaluated in zero-shot and few-shot settings.The Banking77-derived queries may inherit labeling errors, and translationese may mislead model training.Source: §5.4.3 · p. 12

5.4.4. Fact Checking

  • AfrIFact covers 10 African languages, including Wolof, with more than 18,000 claims and annotated evidence spans; evaluations found weak cross-lingual retrieval and fact verification.Few-shot prompting improved results by up to 43% and task-specific fine-tuning added a further 26% gain.Source: §5.4.4 · p. 12

5.4. Text Classification

  • Across four sub-tasks, recent work increasingly uses cross-lingual transfer from multilingual pretrained models, with French generally a more effective pivot than English.Annotated text-classification corpora are absent for Pulaar, Sérère, and the other three national languages, and social-media data bring domain and demographic biases.Source: §5.4.4 · p. 12

5.5. Lexicons & Spell Checking

Wolof spell-checking work ranges from lexical and rule-based methods to neural correction, while shared lexical resources aim to standardize written forms. The reviewed approaches remain constrained by resource availability and limited handling of context and morphology.

Spell-checking approaches

  • Automatic spelling correction identifies misspellings and ranks replacements; reviewed approaches use expert rules, context models, or learned error patterns.The section motivates correction by noting that spelling errors and unnormalized text can reduce readability and informational value.Source: §5.5 · p. 12

Earlier approaches

  • Earlier Wolof spell-checking work did not produce a functional checker, and the Wolof dictionary from the DiLAF project was unavailable online at the time described.The proposed earlier implementation used a French–Wolof dictionary and a Wolof morphological analyzer.Source: §5.5 · p. 13

Spell checking

  • Contextual and morphology-aware correction remains important because Wolof’s concatenated morphemes can make an erroneous suffix look like a valid word to a word-level checker.The reviewed checker combines tries, dynamic programming, and weighted Levenshtein distance, but does not consider context.Source: §5.5 · p. 13

Neural correction

  • Wolof correction research includes synthetic-data LSTM and Transformer models and a 3,400-pair informal-to-standard dataset used to normalize social-media writing.The later model leaves code-switched French and English segments untouched (Sy and Doucouré, 2026a).Source: §5.5 · p. 13

Lexical resources

  • iBaatukaay offered free collaborative lexical data for Wolof, Pulaar, and Bambara, while Senter-mino aims to centralize validated scientific and technical terminology in national languages.Senter-mino is presented as a way to harmonize terms and mitigate code-switching.Source: §5.5 · pp. 13–14

5.6. Machine Translation

Senegalese machine translation has progressed from locally trained systems to multilingual pretrained models and commercial services, but data availability and evaluation quality remain central concerns. Reported local performance gains are associated primarily with corpus size and pretrained backbone quality, while domain-specific parallel data remains a stated bottleneck.

Evaluation and training resources

  • Early MT resources included NTREX-128 with about 2,000 English-to-Wolof evaluation sentences and SMOL, a 6.1-million-token training suite covering Wolof and Pulaar.These initiatives are described as contributing to machine translation for low-resource languages.Source: §5.6 · p. 14

Local MT studies

  • Local Wolof–French MT work includes 70,000- and 83,000-sentence corpora, but earlier evaluations included accuracy-based assessment and substantial train–test overlap.About 60% of the test sentences in one study were also in training; locally created French–Wolof datasets discussed here were not publicly available (Dione et al., 2022).Source: §5.6 · p. 14

Local MT results

  • 26.38 BLEU is the reported French→Wolof result for fine-tuned SMALL-100 trained on 175,000 parallel sentences; the paper attributes progress primarily to data and pretrained-backbone quality, not architecture alone.It identifies domain-specific parallel corpora in health, legal, and educational registers as the remaining practical bottleneck.Source: §5.6 · p. 15

Multilingual models

  • Multilingual MT efforts include African-language models and benchmarks, while fine-tuning SMALL-100 on an expanded Wolof–French corpus yielded the 26.38 BLEU result.The review also describes AfriNLLB as publicly releasing its models and training data, and other multilingual resources including Toucan and AfroLingu-MT.Source: §5.6 · p. 15

Landscape and resources

  • The review groups Senegalese MT into local small-corpus systems, multilingual fine-tuning, and proprietary systems, and identifies domain-specific data and reproducibility as continuing concerns.MudawanSn is an exception to the pattern of unreleased local French–Wolof data: it is an openly released Wolof–Arabic corpus of 1,271 manually translated pairs (Mbaye and Diop, 2026).Source: §5.6 · p. 16

5.7. Question Answering and Dialogue Systems

The reviewed work spans dialogue systems, question answering, and language models, with persistent performance gaps for Wolof and other low-resource African languages. Available datasets, community-led data creation, and multilingual systems provide research pathways, while the paper also describes practical conversational applications and modular-system trade-offs.

Datasets and benchmarks

  • African-language dialogue and QA resources include 1,500 translated turns for Wolof and five other languages, plus AfriQA’s more than 12,000 examples across 10 languages including Wolof.A separate multiple-choice reading-comprehension dataset covers 122 language variants including Wolof.Source: §5.7 · p. 17

Instruction and pretraining data

  • Wolof was added to the AYA instruction dataset through GalsenAI involvement, but insufficient data led to its exclusion from the resulting AYA models; AfriqueLLM also excluded languages below 90 million tokens.The review presents these cases as highlighting the importance of sufficiently large, clean pretraining corpora.Source: §5.7 · p. 17

Models and corpora

  • The passage reports that models trained on language-specific corpora produced by this pipeline outperform models trained on other public web-based multilingual datasets, and describes a synthetic Wolof corpus with 51,000 high-quality documents.FineWeb2-HQ-50k-Wolof was produced by machine-translating a curated subset of FineWeb2 into Wolof using the Oolel model.Source: §5.7 · p. 18

LLMs and performance

  • Wolof is among the lowest-performing languages in LLM evaluations, with the review linking this result to scarce and poor-quality web data and identifying targeted data creation and curation as the main leverage point.The cited estimate is fewer than 50 million publicly available Wolof characters.Source: §5.7 · pp. 16, 18

Applications

  • Senegalese conversational applications include Wolof customer-service and health chatbots, while Sayt’u Hemophilie combines French–Wolof text and voice interaction and received a SUS score of 81.7.The review describes these systems as modular pipelines involving speech recognition, language understanding, and response generation.Source: §5.7 · pp. 18–19

System design

  • Modular dialogue systems are more data-efficient than LLMs but require separate component optimization and can suffer error cascading, while the closed AWA system has an open-source Wolof alternative in Oolel.Oolel is based on Qwen 2.5 and combines current AI technology with Wolof linguistic expertise (Sy and Doucoure, 2024).Source: §5.7 · p. 19

5.8. Speech Processing

Speech processing spans recognition, synthesis, and dialogue, with the reviewed work showing substantial progress for Wolof alongside uneven coverage across Senegalese languages. Community data, correction, and language-specific adaptation recur as important elements of that progress.

5.8.1. Automatic Speech Recognition

  • Synthetic voice data for African ASR could be created for less than 1% of real-data collection costs and may complement human data.Source: §5.8.1 · p. 20
  • 8.56% WER for M-Kiriku on FLEURS-Wolof contrasts with 97.4% for Whisper Large v3, while the review attributes the spread to adaptation, tokenization, and training-data domain.The review concludes that Wolof has a credible open ASR pipeline, while comparable coverage for the other national languages remains a frontier.Source: §5.8.1 · pp. 21–22
  • Wolof ASR developed from HMM systems trained on tens of hours through large-scale self-supervised pretraining to LLM-inspired encoder-decoder systems, with community correction and local fine-tuning delivering strong results.The ALFFA work released data and scripts based on 18,000 utterances and over 21 hours of Wolof speech; GalsenAI later corrected a major WAXAL Wolof-data misalignment.Source: §5.8.1 · pp. 20, 22

5.8.2. Speech Synthesis

  • Wolof TTS has multiple open systems, while Pulaar, Sérère, and the other three national languages remain at single-model proof-of-concept stage without domain or speaker diversity for practical deployment.WAXAL-TTS and africa-g2p are identified as actionable resources for closing this gap.Source: §5.8.2 · pp. 22–23

5.8.3. Spoken Dialog Systems

  • The first Wolof speech language model reused compute through continued pretraining, while an ASR-free keyword-spotting approach underperformed ASR with five minutes of training data in controlled same-domain experiments.Source: §5.8.3 · p. 23
  • Spoken-dialog work moves from modular pipelines limited by cascading errors toward end-to-end and speech-LLM approaches that promise better fluency but require far larger resources.The review identifies keyword-spotting and intent-classification datasets for all six national languages as a shared evaluation foundation for speech understanding.Source: §5.8.3 · p. 23

Cross-Language and Cross-Task Coverage Summary

The resource map shows uneven NLP task coverage across Senegal’s six national languages: Wolof has coverage across all task categories, while the remaining five have full coverage only in keyword spotting.

Cross-Language and Cross-Task Coverage Summary

  • Wolof has some form of resource coverage across all task categories, whereas the other five languages collectively have full coverage only in keyword spotting, the only task with a dataset released simultaneously for all six.Table 2 marks dedicated coverage with a filled circle, minimal or partial coverage with a half-filled marker, and no published resource with a dash.Source: Cross-Language and Cross-Task Coverage… · pp. 23–24

6. Case Study: NLP Pipelines for the Social Sciences

The section describes how NLP could support multilingual qualitative social science research in Senegal, especially by easing transcription, translation, and analysis. It also highlights privacy and methodological challenges that require careful validation and transparent practice.

6. Case Study: NLP Pipelines for the Social Sciences

  • Interviews, focus groups, and participant observations in multiple languages make manual transcription and translation time-consuming, while ASR and machine translation could reduce transcription time and costs.The passage notes that personal data in already-transcribed interviews makes them difficult to collect for ASR training.Source: §6 · p. 24
  • Automating qualitative analysis raises ethical and privacy challenges, and LLM use for empirical research has unresolved reliability, validity, replicability, and model-drift concerns.Suggested practices include transparent model specifications, validation against human-coded ground truth, and explicit handling of nondeterminism.Source: §6 · p. 24
  • For low-resource languages such as Wolof, methodological opacity or unvalidated model behavior could disproportionately distort scientific findings.The discussion presents replicability and careful evaluation as a roadmap for responsible NLP use in linguistic and socio-cultural analyses.Source: §6 · p. 24

7. Discussions & Perspectives

Senegalese NLP faces linked gaps in data, language coverage, infrastructure, skills, visibility, and governance, while community-centered opportunities include typology-aware methods, shared resources, and interdisciplinary collaboration.

7.1. Challenges

  • Across task clusters, high-quality standardized data remain scarce and fragmented, with Wolof better served than Joola, Mandinka, and Soninké.Corpora often lack size, metadata, or representativeness; speech recordings have limited duration and speaker diversity, and annotation is slow.Source: §7.1 · pp. 24–25, Table 2 · p. 25
  • Tonal contrasts, morphophonemic alternations, evolving orthographies, and inconsistent transliteration complicate tokenization and data integration.These challenges can reinforce Wolof’s dominance over other local languages.Source: §7.1 · p. 25
  • Limited local hardware and operational compute infrastructure coincide with shortages of trained researchers, engineers, and experienced mentors.West Africa’s linguistic and cultural expertise availability score is 2.0 out of 5.0; formal university programmes average 3.5 for job readiness, versus 4.5 for innovation hubs and project-based learning.Source: §7.1 · p. 25
  • Local research visibility and access remain constrained; 50% of local Sentiment Analysis papers are behind paywalls, while ethical data governance remains underdeveloped.The paper links the concentration of paywalled work in this active sub-field to constraints on reproducibility and community uptake.Source: §7.1 · pp. 25–26, Figure 8 · p. 26

7.2. Opportunities and Future Directions

  • Typology-aware systems should preserve tone, exploit Wolof vowel harmony, and represent Mandinka serial verb constructions explicitly.The paper warns that stripping tone diacritics can degrade task quality, while standard architectures may mishandle serial verb constructions.Source: §7.2 · p. 26
  • Shared resources and regional partnerships could support multilingual models, improve generalization, and include underrepresented languages in continental benchmarks.The paper identifies collaboration with regional initiatives as one route to interoperability.Source: §7.2 · pp. 26–27
  • A centralized GitHub catalogue supports transparency and reuse, while open data, reproducible tools, and language-specific quality protocols remain priorities.The catalogue gathers datasets, models, and tools across surveyed tasks, with source links and available license information.Source: §7.2 · p. 27
  • Sustainable progress pairs interdisciplinary training and community engagement with ethical governance, while NLP may support more timely and inclusive social-science research.The paper also identifies language literacy as a structural prerequisite for data collection, annotation, error correction, and community validation.Source: §7.2 · pp. 27–28

8. Conclusion

The paper synthesizes NLP initiatives and resources across Senegal’s six national languages in a social-science context. It identifies shared infrastructure, investment, and collaboration as essential to realizing the field’s potential, while leaving dataset evaluation and benchmark implementation for future work.

8. Conclusion

  • The paper presents a comprehensive synthesis of NLP initiatives and resources for Senegal’s six national languages, situated within the social sciences.The languages covered are Wolof, Pulaar, Sérère, Diola, Mandingue, and Soninké.Source: §8 · p. 28
  • NLP integration could enhance social-science research productivity and inclusiveness through automated transcription, translation, and thematic analysis.The conclusion calls for sustained investment in open data, interdisciplinary training, ethical governance, and shared infrastructure.Source: §8 · p. 28
  • Qualitative evaluation of existing datasets and implementation of benchmarks across NLP tasks remain planned future work.The survey’s findings provide a foundation for identifying research priorities and formulating agendas.Source: §8 · p. 28

Related papers

Loading 2601.09716v2…