{"id":"c1272b77-874b-47a3-b13a-bbcdef3f5a6a","arxiv_id":"2411.19017","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey cataloging datasets, features, and machine-learning methods for automatic hate speech detection in low-resource languages, organized by world region, with an overview of open challenges.","lead":"This paper surveys research on automatic hate speech detection in low-resource languages, cataloging datasets, features, and machine-learning techniques across Europe, Africa, Latin America, and Asia. It is a reference work that could help researchers and platform moderators find existing resources and understand gaps in multilingual hate speech detection.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's core claim of comprehensive, up-to-date coverage is not verifiable: the §3.1 search is underspecified and no screening counts are given, and the paper's own §5.2 definition of 'low-resource' includes well-resourced languages such as Spanish and French.","rationale":"The reader's weakest assumption is correct: the survey's value depends on a representative, reproducible literature search, and Section 3.1 is too underspecified to establish that. My concrete test targets exactly this gap. I add a second, related load-bearing concern: the paper's own definition of 'low-resource' (§5.2) includes well-resourced languages, so the survey's scope is internally broader than its title and abstract claim. That is not a disagreement with external consensus; it is an internal inconsistency that affects what the survey can claim. I also note Table 2 row 9 as a concrete reporting error, though it alone would not shift the verdict. I give the paper credit for a clear structure, a broad set of languages, useful dataset tables with links, and discussion of prior surveys; these are genuine assets. The concerns are not fatal to the survey's usefulness, but they do require correction and a more careful scope statement, which is exactly the reader's conditional verdict. Therefore I recommend no change to the reader's verdict.","tokens_in":32859,"tokens_out":3940,"duration_ms":48820,"concrete_test":"Reproduce the §3.1 search exactly as documented: for each keyword in Table 3, run combined queries such as 'hate speech detection' AND each language keyword in IEEE Xplore, ACM DL, and Science Direct for 2017–2024; record total hits, records screened, and records included. Compare the included set against the paper's bibliography plus a hand-curated gold set of 20 recent low-resource hate speech papers (e.g., selected from HASOC, DravidianLangTech, WOAH, and LREC/COLING proceedings). If recall on the gold set is below about 80%, or if reproducing the search requires undocumented judgment calls, the 'detailed survey' claim fails as stated. Separately, verify Table 2 row 9: an independent check should show Religion=Yes and Ableism=No for a religion-targeted statement, contradicting the printed row.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that this is a detailed, current survey of hate speech detection in low-resource languages. For that claim to hold, two conditions must be met: (a) the literature search is systematic enough to support 'detailed' and 'comprehensive', and (b) the 'low-resource' scope is applied consistently. Condition (a) is not met as reported. Section 3.1 lists broad database names (Google Scholar, IEEE Xplore, ACM DL, Science Direct) and a keyword table, but gives no exact query strings, no search dates, no number of records retrieved, no deduplication counts, and no PRISMA-style flow. The filtering step (§3.2) says papers with 'indeterminate information' were excluded but does not define that criterion. Because the contribution is a survey, this undocumented pipeline makes the coverage claim unfalsifiable and unreproducible. Condition (b) is explicitly violated in §5.2: 'except for English, we have considered all the other languages as low-resource languages.' This places Spanish, French, German, Portuguese, Arabic, and Korean inside 'low-resource', diluting the stated focus and changing what the survey's conclusions about 'low-resource languages' actually mean. A further concrete fidelity error appears in Table 2, row 9: the example 'People from <insert religion> should not be allowed inside' is described in the text as religious hate speech, but the table marks Religion as 'No' and Ableism as 'Yes'. This is a small but specific sign that tabular reporting cannot be taken at face value without checking.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of automatic online hate speech detection, with a stated focus on low-resource languages. It reviews definitions of hate speech from major platforms, proposes a categorization with overlapping concepts, describes a keyword-based literature search, catalogs English and non-English datasets (monolingual, multilingual, and multimodal), summarizes detection methods by region (Europe, Latin America, Africa, Asia, and the Indian subcontinent), and closes with research challenges and future directions. The central claim, stated in the abstract, is that the article provides a detailed and current survey of hate speech detection in low-resource languages, including available datasets, features, and techniques.","tokens_in":33132,"tokens_out":3574,"duration_ms":30960,"significance":"If the coverage claims are trustworthy, the survey would be a useful entry point for researchers working on non-English hate speech detection, especially because of its consolidated dataset tables (Tables 4 and 5), its continent-wise organization, and its attention to Indic languages. The paper's main strengths are breadth of languages surveyed and the structured tabulation of many datasets and methods. However, the survey's core value depends on two conditions: that the literature search is systematic enough to support the word 'detailed'/'comprehensive,' and that the term 'low-resource' is applied consistently. The manuscript currently does not meet either condition: the search protocol in Section 3.1 is under-specified, and Section 5.2 explicitly classifies Spanish, French, German, Portuguese, and other relatively well-resourced languages as low-resource. There is also a concrete fidelity error in Table 2. The paper provides no machine-checked proofs, reproducible code, or parameter-free derivations; its contribution is a narrative synthesis of the secondary literature.","major_comments":[{"comment":"The literature search is not described in enough detail to support the claim of a comprehensive survey. The manuscript lists broad database names and a table of keyword categories, but gives no exact query strings, no search dates, no number of records retrieved per database, no deduplication counts, and no screening counts. The exclusion criterion in Section 3.2, 'Research papers with indeterminate information were excluded,' is not defined. Because the contribution is explicitly a survey, this undocumented pipeline makes the coverage claim unfalsifiable and unreproducible. The authors should add a full search log, inclusion/exclusion criteria, and a flow diagram, or substantially soften the comprehensiveness claim to that of a selective review.","section":"Section 3.1 (3.1.1 and 3.1.2)"},{"comment":"The definition of 'low-resource' is internally inconsistent: 'except for English, we have considered all the other languages as low-resource languages' includes Spanish, French, German, Portuguese, Arabic, and Korean, several of which are among the most resourced non-English languages. This contradicts the paper's own language-distribution charts (Fig. 4), where Spanish, German, and Arabic are shown as the second-most-used languages on major platforms. The broadened scope dilutes the stated focus and changes what the survey's conclusions about 'low-resource languages' actually mean. The authors should adopt an operational definition (e.g., based on dataset size, NLP tool coverage, or speaker population) and apply it consistently throughout.","section":"Section 5.2"},{"comment":"The example 'People from <insert religion> should not be allowed inside.' is described in the text as religious hate speech that is neither abusive nor cyberbullying, but the table marks Religion as 'No' and Ableism as 'Yes'. This is a direct contradiction between the table and the text. The row must be corrected, and the remaining rows should be re-checked for similar inconsistencies.","section":"Table 2, row 9"}],"minor_comments":[{"comment":"The abstract ends with 'Keywords: article, template, simple', which appears to be a leftover template string and should be removed before submission.","section":"Abstract"},{"comment":"The label distribution sums incorrectly: 344 invalid comments out of 6375 is approximately 5.4%, not 0.05% as printed. Please correct the percentage and verify the arithmetic in the surrounding dataset descriptions.","section":"Section 4.2.1, Tulkens dataset"},{"comment":"The phrase 'Research papers with indeterminate information were excluded' is vague; please specify what information was considered indeterminate and how this determination was made.","section":"Section 3.2"},{"comment":"The sentence 'The Word2Vec model with Naïve Bayes classifier achieved the best performance with accuracy and ROC Score of 79.83% and 0.8305, respectively' is clear, but the preceding sentence about the Apache Spark-based model would benefit from a citation to the specific model description or the paper being summarized.","section":"Section 5.2.3, Amharic"},{"comment":"The description of Kim et al. [82] refers to 'racist hate speech' measured with 'adoption thresholds of racist hate speech,' but the paper's focus on Korean racism may be clearer if the target group were named explicitly; please check that the summary matches the cited paper.","section":"Section 5.2.4, Japanese subsection"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an early preprint: the abstract contains a leftover template keyword string, some dataset statistics are inconsistent, and the search protocol is not documented in a way that would let a reader reproduce the survey. These issues are fixable. I recommend major revision rather than rejection because the underlying content—the dataset catalog and the language-by-language method summaries—is potentially useful. The two load-bearing problems are the reproducibility of the literature search and the inconsistent definition of 'low-resource'; both must be addressed before the paper can be accepted as a comprehensive survey. If the authors can add a transparent search methodology and tighten the scope, the paper could become a solid reference; otherwise, the claims of comprehensiveness should be substantially reduced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serviceable, up-to-date survey of hate speech detection in low-resource languages, with a genuinely useful continent-wise catalog of datasets and methods. It does not do anything new methodologically, but as a reference map it beats asking newcomers to dig through a decade of shared tasks. The soft spots are all fixable, but they are real.\n\nWhat's good: the paper updates earlier surveys (Fortuna & Nunes, Jahan & Oussalah) with newer resources like SOLD-2024 and Austrotox, and it organizes the work by region, which is helpful for people trying to find work on, say, Dravidian or African languages. The dataset tables give size, labels, and links. The digest of techniques per language is a reasonable entry point. I checked a handful of cited numbers against the sources I know; they're broadly accurate.\n\nWhere it gets wobbly: the survey's stated goal is 'detailed' and 'comprehensive,' but Section 3.1 does not report the actual search. No query strings, no search dates, no counts of retrieved or screened records, and no definition of 'indeterminate information' in Section 3.2. For a survey, that makes the coverage claim unverifiable. Also, Section 5.2 says 'except for English, we have considered all the other languages as low-resource languages.' That puts Spanish, French, German, Portuguese, Korean, and Arabic under the 'low-resource' umbrella. That's not a defensible use of the term, and it blurs what the survey's conclusions mean.\n\nThere are also small fidelity errors. Table 2, row 9 marks a religion-based statement as 'No' for Religion and 'Yes' for Ableism, while the text says it's religious hate speech. The abstract has a leftover template keyword line. These are minor, but they add to the sense that the manuscript wasn't carefully checked.\n\nNet: the central organizational and descriptive work holds up. The paper is a decent entry point for a newcomer, not a definitive reference. I would send it to peer review, but with major revision requests: report the search pipeline in a reproducible way, fix the 'low-resource' definition, and fix the table/text inconsistencies.","headline":"Useful updated map of datasets and techniques for low-resource hate speech detection, but the coverage claim is unverifiable as reported and the scope is overbroad.","tokens_in":33716,"tokens_out":1799,"would_cite":false,"duration_ms":17408,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Survey maps the state of hate speech detection in low-resource languages, cataloging datasets and methods.","keywords":["hate speech detection","low-resource languages","dataset survey","natural language processing","deep learning","multilingual detection","Indic languages","social media"],"falsifier":"Replicating the Section 3.1 search with the listed keywords in the named digital libraries and checking whether the catalog omits a substantial body of published low-resource-language datasets or detection studies, particularly from regional venues, would settle whether the survey's coverage claim holds.","tokens_in":32639,"feed_emoji":"🗺️","tokens_out":5261,"duration_ms":45666,"temperature":0.7,"pith_summary":"This paper is a survey rather than a new detection method. Its goal is to establish a current, structured picture of automatic hate speech detection for low-resource languages: which datasets exist, which features and machine- or deep-learning techniques have been tried, and where the research gaps are. The authors organize the literature by continent and language family, give special attention to Indic languages, and draw a trajectory from TF-IDF and classical classifiers toward word embeddings, LSTM variants, and transformer models. If the survey is faithful, it gives researchers an entry point that previously had to be assembled from many separate papers.","feed_headline":"Survey maps hate-speech datasets in low-resource languages","feed_subtitle":"Catalog of datasets, features, and models shows the field is sparse but moving to transformers.","key_machinery":"The load-bearing device is the survey's taxonomy, built through a keyword-driven search and manual filtering. Datasets are sorted into English, monolingual low-resource, multilingual, and multimodal buckets; detection studies are sorted by continent and, for Asia, into Indic and non-Indic languages. This structure lets the paper convert a collection of individual papers into comparative claims about resource availability, dominant methods, and open problems.","core_discovery":"The paper's central claim is that hate speech detection research is overwhelmingly English-centric and that the work on low-resource languages, though growing, is scattered and uneven. It presents a catalog of monolingual, multilingual, and multimodal hate speech datasets from European, Latin American, African, and Asian languages, together with the feature representations and models applied to them. It also claims that the field has moved from TF-IDF and conventional machine learning to embeddings combined with deep models, with multilingual transformers such as BERT variants and XLM-RoBERTa as the current standard. The survey's accompanying thesis is that the main obstacles are not only scarce data but also culturally specific meaning, ambiguous annotation, biased collection, and the near absence of public code.","pith_inferences":["The continent-wise organization makes explicit that 'low-resource' covers very unequal situations: some European languages resemble English in data availability, while many African and Southeast Asian languages have only a few small datasets; a resource-ranking study could quantify this.","The paper's emphasis on code-mixed and transliterated Indic text suggests a testable extension: models trained on code-mixed Roman-script data may transfer poorly to native-script text unless both scripts appear in training.","Because the survey relies on English-indexed digital libraries, its coverage likely underrepresents work published in regional venues or local languages; searching regional databases would test this gap.","The catalog could be turned into a living benchmark: standardizing annotation labels across datasets would let researchers compare cross-lingual transfer results more reliably."],"forward_implications":["Researchers working on a low-resource language can use the dataset tables to locate existing corpora and shared-task benchmarks before building new ones.","The reported shift from TF-IDF and classical machine learning to word embeddings and fine-tuned transformers implies that new work should start with pre-trained multilingual or monolingual models.","For Indic languages, shared tasks such as HASOC provide reusable evaluation infrastructure, so progress can be measured against common baselines.","The challenges the survey lists imply that better detection requires culturally informed annotation and dataset documentation, not just larger models.","Multimodal hate speech, especially memes and code-mixed text, is an open area where the paper predicts further growth."],"supporting_citations":[{"why":"Baseline systematic review of automatic hate speech detection in text that the survey positions itself as extending.","marker":"[47]"},{"why":"Systematic review of multilingual datasets and deep learning models that underlies the low-resource overview.","marker":"[65]"},{"why":"Survey of hate speech detection in Asian languages, a direct precursor for the Asian and Indic sections.","marker":"[39]"},{"why":"Review of cyberbullying detection methods in low-resource languages that supports the challenges discussion.","marker":"[90]"},{"why":"Multilingual survey including Arabic hate speech, used as a reference for the multilingual facet.","marker":"[4]"},{"why":"English Twitter hate speech dataset that serves as an early benchmark and comparison point.","marker":"[154]"},{"why":"English hate versus offensive dataset that anchors the survey's discussion of annotation ambiguity.","marker":"[35]"},{"why":"HatEval multilingual shared-task dataset that provides benchmark structure for Spanish and English hate speech detection.","marker":"[16]"},{"why":"HASOC-2019 shared task dataset for Hindi and German, foundational for many Indic-language experiments.","marker":"[91]"}],"fun_headline_variants":["Survey: Hate speech detection sparse for low-resource languages","Low-resource languages: hate speech AI's weak spot","Hate speech detection gap widens for low-resource languages","Survey maps hate speech datasets across low-resource languages","Hate speech research: low-resource languages still underserved"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's usefulness depends on its literature search having found a representative sample of the work on hate speech detection in low-resource languages, since the exact queries, databases, and screening rules are only partially documented in Section 3.1.","fun_headline_variants_meta":{"raw":{"variants":["Survey: Hate speech detection sparse for low-resource languages","Low-resource languages: hate speech AI's weak spot","Hate speech detection gap widens for low-resource languages","Survey maps hate speech datasets across low-resource languages","Hate speech research: low-resource languages still underserved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1557,"prompt_tokens":836,"completion_tokens":721,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":640}},"tokens_in":452,"tokens_out":721,"duration_ms":8967,"temperature":1.0,"reasoning_tokens":640,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:36:53.639656+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replicating the Section 3.1 search with the listed keywords in the named digital libraries and checking whether the catalog omits a substantial body of published low-resource-language datasets or detection studies, particularly from regional venues, would settle whether the survey's coverage claim holds.","supporting_citations":[{"cited_title":"Hateful symbols or hateful people? predictive features for hate speech detection on twitter","cited_arxiv_id":null,"evidence_quote":"English Twitter hate speech dataset that serves as an early benchmark and comparison point."}],"review_version":1}