CongoNames Corpus: A Large-Scale Labeled Dataset of Congolese Personal Names

Abstract

Personal names carry cultural and linguistic identity, yet most African countries lack large-scale, structured name datasets suitable for natural language processing (NLP) research and computational social science. We present CongoNames, the first large-scale corpus of personal names from the Democratic Republic of the Congo (DRC), derived from publicly released national secondary-school examination palmarès (result lists) 8,053,983 published annually by the DRC Ministry of Education. The corpus comprises name records spanning 16 examination years (2008–2023) across 12 provinces and 304 sub-provincial regions, each enriched with a reported sex marker (M/F) and regional provenance metadata. We describe a fully deterministic, layered processing pipeline (bronze–silver–gold architecture) that converts raw protable document format (PDF) documents into structured comma-separated values (CSV) datasets without manual annotation or machine-learning-based inference. The dataset is validated against school-level census counts extracted from the same source PDFs, yielding extraction error rates below 2% for all years except 2023 (7.81%, flagged due to a layout change). Descriptive analyses document name length and token-count distributions, character-level n-gram profiles, provincial diversity indices, and inter-provincial name-inventory overlap, collectively establishing the dual linguistic origin—locally rooted Bantu components and Christian/French-origin components—that characterize modern Congolese naming practice. The dataset, processing code, and documentation are released openly to support research in African natural language processing (NLP), onomastics, and computational social science.

Dataset: https://doi.org/10.5281/zenodo.19809985.

Dataset License: Creative Commons Attribution 4.0 International

1. Summary

Personal names are a fundamental carrier of cultural and linguistic identity, and large-scale name corpora support a wide range of applications, including named entity recognition (NER) [1], sociolinguistics, and language-technology personalization [2,3]. Despite the growing momentum of African NLP research [4,5,6], African countries remain substantially underrepresented in widely used name datasets and global language resources [7]. As Ngũgĩ wa Thiong’o observes, treating African names as errors are not merely a technical issue but a cultural one [8]—a perspective that motivates the construction of high-quality, auditable African name resources.

The Democratic Republic of the Congo (DRC) presents a particularly important case. It is one of the most linguistically diverse countries in the world, with hundreds of spoken languages [9]; its naming practices reflect this diversity by drawing simultaneously on indigenous languages (Lingala, Swahili, Kikongo, Tshiluba, and many others), French, and religious traditions. Despite this rich onomastic landscape, no large public corpus of Congolese personal names has previously been available.

To address this gap, we introduce the CongoNames Corpus, a large-scale dataset of Congolese personal names. The corpus is derived from the Examen d’État national secondary-school examination result lists (palmarès), publicly released by the DRC Ministry of Education by province and year. These records provide an authoritative, high-coverage view of names spanning 16 years, with sex-marker and regional-provenance metadata obtained directly from the administrative source without additional annotation. The released dataset does not contain explicit ethnicity labels; province and year should be interpreted as regional provenance, not as ethnic markers.

The main contributions of this paper are as follows:

• We compile 8,053,983 personal-name records from the DRC, each linked to a reported sex marker (M/F) and regional provenance (province and year).
• We present and release a fully reproducible, layered processing pipeline (bronze–silver–gold) that downloads palmarès PDFs, parses and normalizes the extracted text with deterministic rules, and exports structured CSV datasets.
• We provide descriptive statistics and analyses—including distributions by region, year, and sex marker; character-level n-gram profiles; diversity indices and inter-provincial Jaccard similarity—illustrating regional naming variation and the dual linguistic origin of Congolese personal names.
• We validate the extraction pipeline against school-level census counts embedded in the source PDFs, providing quantitative coverage estimates for each examination year.

The remainder of this paper is organized as follows: Section 2 describes the dataset content, structure, and statistical properties; Section 3 details the data collection, processing pipeline, and quality assurance procedures; and Section 4 provides usage guidance, applications, limitations, and ethical considerations.

2. Data Description

2.1. Dataset Overview and Structure

Table 1. CongoNames dataset overview.

Metric	Value
Total records	8,053,983
Unique names (exact)	6,371,941
Unique names (normalized)	6,370,887
Unique regions	304
Unique provinces	12
Unique years	16
Unique source files	469
Simple names	6,324,763
Complex names	1,729,220

provides a high-level summary of the corpus. The dataset comprises 8,053,983 total records drawn from 469 source PDF files spanning 16 examination years (2008–2023) and 304 distinct regions grouped into 12 provinces. The unit of observation is a candidate-level full personal-name record. Accordingly, the total of 8,053,983 refers to complete name strings associated with candidate records rather than individually counted first-name, middle-name, or surname tokens. The near-identical numbers of exact and normalized unique names (6,371,941 vs. 6,370,887) indicate that normalization merged only a small number of near-duplicate variants. This result suggests that the corpus reflects substantial genuine lexical diversity rather than variation introduced by inconsistent formatting or spelling. Exact uniqueness is computed over complete extracted name strings, whereas normalized uniqueness is computed after deterministic case and spacing normalization. No phonetic or language-specific variant merging is applied, for example, distinct spellings such as Ngoy, Ngoie, and Ngoyi remain separate names unless they are identical after the released normalization rules. Approximately 78.5% of records (6,324,763) follow a three-token naming structure (simple names), while the remaining 1,729,220 records (complex names) follow non-three-token structures, including two-token records and records with four or more tokens.

2.2. Released Files and Schema

The corpus is released as multiple CSV files stored under dataset/gold/. All name strings are normalized to lowercase.

Table A1. Schemas of the core candidate-level datasets released with CongoNames. The names_featured file is the recommended starting point for most analytical use cases. (a) names. (b) names_featured.

(a)
Column	Type	Description
id	string	Candidate identifier.
name	string	Full candidate name (lowercase).
sex	string	Reported sex marker (m or f).
year	string	Examination year parsed from filename.
region	string	Region label parsed from filename.
filename	string	Source gold-layer text filename.
line	integer	Line number in gold-layer text file.
(b)
Column	Type	Description
id	string	Candidate identifier (left-padded).
name	string	Full candidate name (lowercase).
sex	string	Reported sex marker (m or f).
year	string	Examination year parsed from filename.
region	string	Region label parsed from filename.
filename	string	Source gold-layer text filename.
line	integer	Line number in gold-layer text file.
words	integer	Number of whitespace tokens ($\left|N\right|$).
length	integer	Character count including spaces ($\ell \left(N\right)$).
category	string	simple (3 tokens) or complex.
province	string	Province inferred from region mapping.

documents the two primary candidate-level files. The names file is the core release artifact, providing one row per candidate with a unique identifier, the full name string, the reported sex marker, the examination year and region parsed from the source filename, the source filename, and a gold-layer line number for full provenance traceability. The names_featured file extends names with four computed fields: word count ($\left|N\right|$), character length ($\ell \left(N\right)$, including spaces), structural category (simple or complex, i.e., $C\left(N\right)$), and an inferred province label derived via a deterministic region-to-province mapping. This file is the recommended starting point for most analytical use cases.

Listing 1 presents an illustrative JavaScript Object Notation (JSON) schema representation of a synthetic candidate record and its associated feature extraction and name decomposition. A synthetic candidate with the full name Ilunga Ngoy Jean would appear in names.csv as a single record, with name = Ilunga Ngoy Jean, sex = m, year = 2015, region = Katanga, together with the corresponding source filename and line number. In names_featured.csv, the same record would have words = 3, length = 16, category = simple, and province = katanga. In names_components.csv, the name would be decomposed into three token records—Ilunga, Ngoy, and Jean. For the exploratory decomposition used in this study, Ilunga Ngoy would be treated as a transliterated example of the probable locally rooted component $G\left(N\right)$, whereas Jean would represent the probable Christian/French-origin component $S\left(N\right)$. This example is purely illustrative and does not correspond to an actual candidate record.

Listing 1: Illustrative JSON representation of a synthetic candidate full name record.

1 { 2   “name”:“Ilunga Ngoy Jean”, 3   “sex”:“m”, 4   “year”:2015, 5   “region”:“katanga”, 6   “features”:{“words”:3,“length”:16, 7          “category”:“simple”, 8          “province”:“Katanga”}, 9   “components”:[“Ilunga”,“Ngoy”,“Jean”], 10   “decomposition”:{“probable_native”:“Ilunga Ngoy”, 11             “probable_surname”:“Jean”} 12 }

2.3. Name Structure and Basic Statistics

Table 2. Name structure summary.

Metric	Value
Total records	8,053,983
Mean characters per name	20.19
Median characters per name	21
90th percentile characters	25
95th percentile characters	26
Mean tokens per name	2.86
Median tokens per name	3
90th percentile tokens	3
95th percentile tokens	3

reports structural summary statistics for the full corpus. The mean character length per name is 20.19, with a median of 21 and a 95th-percentile value of 26, indicating a compact and consistent length distribution with little extreme variation. The mean token count is 2.86, and both the 90th and 95th percentiles equal 3, confirming that the three-token structure is strongly dominant. These figures provide practical guidance for downstream modeling, for example, when setting maximum sequence lengths for character-level or subword models.

To illustrate token-role regularities arising from the heuristic decomposition into a probable native-name component $G\left(N\right)$ (the first two tokens of a simple name) and a probable surname component $S\left(N\right)$ (the final token), Table 3 lists the ten most frequent first and last tokens. The most frequent first tokens of $G\left(N\right)$—Ilunga (75,206), Ngoy (58,229), and Kasongo (53,821)—are recognizably Bantu names associated with the Katanga and Kasai regions, reflecting the geographic weight of those provinces. In contrast, the most frequent final tokens of $S\left(N\right)$—Jean (117,477), Joseph (51,895), Marie (50,889), and Esther (49,688)—are overwhelmingly Christian or biblical in origin. This asymmetry is one of the clearest empirical signals in the corpus of the systematic co-occurrence of locally rooted and Christian-origin name elements that characterizes modern Congolese naming practice. These patterns should be interpreted cautiously: $G\left(N\right)$ and $S\left(N\right)$ are heuristic components and do not constitute ground-truth annotations.

2.4. Distribution by Province and Year

Table 4. Name distribution by province.

Province	Total	Unique	M	F	Simple	Complex
Autre	2,522,510	2,465,765	1,530,087	992,423	2,018,242	504,268
Bandundu	801,752	775,347	496,382	305,370	293,673	508,079
Bas-Congo	292,101	288,681	169,699	122,402	269,494	22,607
Équateur	373,646	372,464	265,121	108,525	332,519	41,127
Kasai-Occidental	437,913	411,429	320,103	117,810	322,189	115,724
Kasai-Oriental	397,775	383,110	277,774	120,001	369,505	28,270
Katanga	846,746	818,704	543,810	302,936	698,197	148,549
Kinshasa	1,165,038	1,142,275	570,529	594,509	1,091,330	73,708
Maniema	129,114	123,260	97,669	31,445	70,018	59,096
Nord-Kivu	392,793	386,407	224,967	167,826	351,686	41,107
Orientale	340,152	338,730	216,564	123,588	293,619	46,533
Sud-Kivu	354,443	343,664	222,041	132,402	214,291	140,152

reports total, unique, sex-disaggregated, and structural counts by province. The “Autre” (other/unclassified) category, Katanga, and Kinshasa account for the largest absolute record counts, reflecting population size and examination participation rates in those areas. Kinshasa is notable for the near-parity of male (570,529) and female (594,509) records—in contrast to other provinces where male records consistently exceed female records.

Table 5. Name distribution by year.

Year	Total	Unique	M	F	Simple	Complex
2008	183,076	164,012	115,278	67,798	168,344	14,732
2009	256,212	255,294	171,348	84,864	196,286	59,926
2010	281,536	280,370	190,608	90,928	211,094	70,442
2011	323,906	322,315	216,895	107,011	233,558	90,348
2012	324,890	319,468	217,107	107,783	246,949	77,941
2013	270,705	269,618	176,261	94,444	209,403	61,302
2014	330,063	328,490	211,911	118,152	254,931	75,132
2015	729,829	365,305	456,623	273,206	559,523	170,306
2016	400,139	398,523	246,542	153,597	318,959	81,180
2017	413,563	412,049	252,718	160,845	330,470	83,093
2018	501,969	499,610	303,193	198,776	406,208	95,761
2019	516,036	513,885	311,569	204,467	422,718	93,318
2020	633,454	629,642	378,949	254,505	521,790	111,664
2021	956,994	492,166	563,006	393,988	790,433	166,561
2022	1,222,198	602,653	713,479	508,719	982,911	239,287
2023	709,413	691,221	409,259	300,154	471,186	238,227

provides the same breakdown by examination year. Both total and unique name counts grow substantially across the 2008–2023 window, reflecting national expansion of secondary-school examination participation. The lower unique-to-total ratios visible in 2015, 2021, and 2022 should, therefore, be read as year-level uniqueness anomalies, not as a change in the counting unit. Totals still count candidate rows, while the unique column counts distinct normalized full-name strings within that year. These years correspond to especially large cohorts in the released tables, so repeated full-name strings, homonymy, broader coverage, and possible residual duplicate extraction have a larger visible effect on the uniqueness ratio. Their aggregate extraction errors remain small (0.07% in 2015, $-0.05$% in 2021, and $-1.76$% in 2022; Section 3.4.2), indicating that the candidate-row totals are broadly consistent with the school-level census checks, although uniqueness-based analyses should treat these three years with additional caution.

Table 3. Most frequent name tokens by position. (a) First tokens ($G\left(N\right)$). (b) Last tokens ($S\left(N\right)$).

(a)
Token	Count
Ilunga	75,206
Ngoy	58,229
Kasongo	53,821
Ntumba	37,248
Kambale	35,852
Mbuyi	35,662
Kavira	34,841
Kasereka	25,418
Kazadi	25,305
Banza	24,864
(b)
Token	Count
Jean	117,477
Joseph	51,895
Marie	50,889
Esther	49,688
Grace	43,163
Pierre	42,834
Dorcas	41,919
David	41,599
Moïse	41,202
Ruth	40,948

Table 3. Most frequent name tokens by position. (a) First tokens ($G\left(N\right)$). (b) Last tokens ($S\left(N\right)$).

(a)
Token	Count
Ilunga	75,206
Ngoy	58,229
Kasongo	53,821
Ntumba	37,248
Kambale	35,852
Mbuyi	35,662
Kavira	34,841
Kasereka	25,418
Kazadi	25,305
Banza	24,864
(b)
Token	Count
Jean	117,477
Joseph	51,895
Marie	50,889
Esther	49,688
Grace	43,163
Pierre	42,834
Dorcas	41,919
David	41,599
Moïse	41,202
Ruth	40,948

2.5. Character-Level Distributions

Figure 1 compares letter-frequency profiles computed over (a) the probable native-name component $G\left(N\right)$ and (b) the probable surname component $S\left(N\right)$, restricted to simple names. The two panels reveal strikingly different distributional signatures. In $G\left(N\right)$, frequencies are spread more evenly across the alphabet, with relatively high shares for consonant clusters characteristic of Bantu phonology (k, m, n, b). In $S\left(N\right)$, the distribution is dominated by letters common in French and Latin-origin words (a, e, i, r, n), consistent with frequently occurring Christian names such as Jean, Marie, and Christine.

At a finer granularity,

Table 6. Top character n-grams (native segments). (a) Bigrams (top 10). (b) Trigrams (top 10). (c) 4-grams (top 10).

(a)
N-Gram	Count	Share
ng	2,690,439	0.0356
ka	2,422,068	0.0321
mb	2,344,781	0.0310
an	2,264,842	0.0300
ba	1,961,024	0.0260
am	1,958,668	0.0259
ma	1,639,772	0.0217
nd	1,279,949	0.0169
mu	1,276,863	0.0169
al	1,263,542	0.0167
(b)
N-Gram	Count	Share
nga	1,076,984	0.0156
mba	976,231	0.0141
ang	785,278	0.0113
amb	768,401	0.0111
ngo	723,730	0.0105
aka	593,891	0.0086
ong	593,060	0.0086
ala	483,541	0.0070
nda	460,683	0.0067
mbo	433,059	0.0063
(c)
N-Gram	Count	Share
amba	399,561	0.0064
anga	379,454	0.0060
ongo	354,151	0.0056
tshi	306,376	0.0049
umba	233,535	0.0037
unga	192,424	0.0031
ombo	189,248	0.0030
anda	168,117	0.0027
enga	151,186	0.0024
ngam	149,237	0.0024

and 

Table 7. Top character n-grams (surname segments). (a) Bigrams (top 10). (b) Trigrams (top 10). (c) 4-grams (top 10).

(a)
N-Gram	Count	Share
an	961,970	0.0286
in	798,644	0.0237
el	777,156	0.0231
ie	754,962	0.0224
ne	715,335	0.0212
er	687,128	0.0204
ri	679,318	0.0202
en	505,480	0.0150
ar	502,520	0.0149
is	495,471	0.0147
(b)
N-Gram	Count	Share
ine	364,959	0.0133
ris	190,542	0.0070
ean	183,197	0.0067
jea	182,725	0.0067
tin	181,496	0.0066
sti	174,469	0.0064
ric	165,581	0.0061
mar	163,844	0.0060
cha	160,033	0.0058
anc	152,924	0.0056
(c)
N-Gram	Count	Share
jean	180,585	0.0086
rist	127,322	0.0061
chri	123,746	0.0059
hris	121,488	0.0058
stin	120,434	0.0057
ette	102,192	0.0049
ranc	91,957	0.0044
line	83,398	0.0040
usti	82,396	0.0039
bert	78,362	0.0037

report the most frequent character n-grams (bigrams, trigrams, and 4-grams) for $G\left(N\right)$ and $S\left(N\right)$, respectively. For $G\left(N\right)$, the leading bigrams—ng (0.0356), ka (0.0321), and mb (0.0310)—together with trigrams nga, mba, and amb and 4-grams amba, are hallmark sequences of Bantu phonology. For $S\left(N\right)$, the leading n-grams (an, in, el; trigrams ine, ris, ean; 4-grams jean, rist, chri) are characteristic of French and Christian-origin names. The complementarity of these two tables provides quantitative, reproducible evidence for the dual-origin structure of Congolese personal names.

2.6. Name Length and Token Count Distributions

Figure 2 shows the character-length distribution $\ell \left(N\right)$ across all records on a log-frequency scale. The distribution is approximately unimodal and right-skewed, peaking in the 18–22 character range and tapering sharply below 10 and above 30. Figure 3 depicts the token-count distribution $\left|N\right|$ on the same scale. The distribution is highly concentrated at $\left|N\right|=3$, with a secondary mode at $\left|N\right|=2$ and a rapidly declining tail for $\left|N\right|\ge 4$. Together, these figures establish the expected input characteristics for downstream models processing CongoNames data.

2.7. Reported Sex Marker and Regional Composition

Figure 4 presents the distribution of reported sex markers (M/F) across the 12 provinces on a logarithmic count scale. Male candidates consistently outnumber female candidates in every province except Kinshasa, with the relative gap varying across regions. Figure 5 shows the same breakdown by examination year. Both counts grow substantially over 2008–2023, and the relative M/F gap narrows modestly in recent years, tentatively suggesting incremental progress in gender parity in secondary-school completion, though rigorous causal analysis is beyond the scope of this data descriptor.

2.8. Name Diversity Indices

We quantify name diversity within each group (year or province) using two standard measures. Let $\mathcal{V}$ be the set of distinct full names in a group and let $p\left(v\right)$ denote the empirical frequency of name $v\in \mathcal{V}$. The Shannon [10] entropy is

$$H=-\sum _{v\in \mathcal{V}}p\left(v\right)logp\left(v\right),$$

(1)

and the Simpson [11] (Gini–Simpson) diversity index is

$$D=1-\sum _{v\in \mathcal{V}}p{\left(v\right)}^2.$$

(2)

We also report the effective number of names (Hill number of order 1 [12]),

$$N_{\mathrm{eff}}=exp\left(H\right),$$

(3)

interpretable as the number of equally frequent names that would yield the same entropy value.

Table 8. Name diversity by year.

Year	Shannon	Simpson	Effective Names
2008	11.972	0.9999934	158,221
2009	12.448	0.9999961	254,823
2010	12.542	0.9999964	279,754
2011	12.681	0.9999969	321,466
2012	12.667	0.9999968	317,185
2013	12.502	0.9999963	268,971
2014	12.699	0.9999969	327,512
2015	12.805	0.9999972	363,981
2016	12.894	0.9999975	397,729
2017	12.927	0.9999976	411,347
2018	13.119	0.9999980	498,471
2019	13.148	0.9999980	512,874
2020	13.350	0.9999984	627,808
2021	13.094	0.9999979	485,789
2022	13.302	0.9999983	598,601
2023	13.424	0.9999983	675,847

and 

Table 9. Name diversity by province.

Province	Shannon	Simpson	Effective Names
Autre	14.705	0.9999996	2,434,210
Bandundu	13.541	0.9999986	760,019
Bas-Congo	12.565	0.9999964	286,366
Équateur	12.826	0.9999973	371,927
Kasai-Occidental	12.868	0.9999966	387,585
Kasai-Oriental	12.831	0.9999971	373,673
Katanga	13.590	0.9999986	798,293
Kinshasa	13.940	0.9999991	1,133,022
Maniema	11.688	0.9999907	119,087
Nord-Kivu	12.856	0.9999973	383,186
Orientale	12.731	0.9999970	338,058
Sud-Kivu	12.725	0.9999968	336,043

report these values by year and by province, respectively. The Simpson index values are uniformly close to 1.0, confirming that no single name dominates any provincial or annual inventory. Across years, both H and $N_{\mathrm{eff}}$ increase monotonically (with minor exceptions), reflecting growing corpus coverage and/or increasing naming variation across cohorts. Across provinces, Maniema records the lowest Shannon entropy (11.688) and effective name count (119,087), consistent with its smaller population and more linguistically homogeneous profile, while Kinshasa and the “Autre” category exhibit the highest diversity values ($H=13.940$ and $H=14.705$, respectively), plausibly reflecting cosmopolitan, multi-ethnic compositions.

2.9. Inter-Provincial Name Overlap

To characterize the similarity of provincial name inventories, we compute pairwise Jaccard similarity over sets of unique full names. For provinces a and b with distinct name sets ${\mathcal{V}}_a$ and ${\mathcal{V}}_b$, the Jaccard [13] similarity is

$$J(a,b)={\displaystyle \frac{|{\mathcal{V}}_a\cap {\mathcal{V}}_b|}{|{\mathcal{V}}_a\cup {\mathcal{V}}_b|}}.$$

(4)

Figure 6 visualizes the resulting similarity matrix. Inter-provincial Jaccard similarities are uniformly low, with most off-diagonal values below 0.05, confirming that provincial name inventories are highly distinctive. The highest pairwise overlaps occur between geographically adjacent or linguistically related provinces (e.g., Kasai-Occidental and Kasai-Oriental; Nord-Kivu and Sud-Kivu), consistent with shared language communities straddling provincial boundaries. Kinshasa shows moderately elevated overlap with several other provinces, plausibly reflecting internal migration from across the country. Maniema consistently records the lowest overlap, in line with its relatively small and linguistically distinct inventory. These patterns accord with the known linguistic geography of the DRC, though caution is warranted given that province is only a coarse proxy for linguistic community.

3. Methods

3.1. Data Source

The CongoNames corpus is derived from the Examen d’État, the national secondary-school graduation examination of the DRC. Each year, the Ministry of Primary, Secondary and Technical Education (MINEPST) publishes the palmarès (result lists) of exam candidates organized by province (historically referred to as education districts), distributed as PDF documents on the official ministry website. Each PDF typically contains candidate names together with school affiliation, reported sex marker, and examination outcome.

Using these publicly available documents as the data source offers three key advantages: (1) they provide an authoritative, high-coverage view of names for each year covered; (2) they include a sex marker (M/F) and a province association, supplying metadata without additional manual annotation and (3) the annual cadence supports temporal analyses across cohorts.

We crawled the ministry’s results portal1 to retrieve all available PDFs. The current release covers 2008–2023 (the most recent year available at the time of extraction, October 2024) and spans multiple provinces. In total, the collection comprises 469 PDFs yielding 8,053,983 individual name entries (Table 1).

All scripts and mapping tables used for downloading, parsing, and normalization are released with the code repository to support end-to-end reproducibility.2

3.2. Pipeline Architecture

Transforming raw PDF files into a structured corpus required a multi-stage pipeline. We adopted a layered data processing approach inspired by the medallion architecture [14], organizing work into distinct bronze, silver, and gold data layers to ensure traceability and incremental refinement. Figure 7 illustrates the end-to-end workflow.

The principal parsing challenges were practical rather than statistical. The ministry published PDF files contain repeated page headers and footers, wrapped candidate lines, school-summary blocks interleaved with candidate lists, inconsistent separators around sex markers and examination outcomes, historical spelling variants in province and region labels, accent and apostrophe variation in French and local names, and a noticeable layout change in 2023. The pipeline addresses these issues through layer-specific preservation, canonical filename mapping, deterministic regular expressions, line-merging rules, and validation against school-level counts rather than through manual annotation or model-based inference.

3.2.1. Bronze Layer: Source Collection and Text Extraction

PDF result lists are automatically collected from the official ministry portal via a web crawler that enumerates province–year result pages, extracts PDF URLs, and downloads each document with retry logic and checksum verification. The raw PDFs are stored unchanged in dataset/bronze/pdf. Each PDF is then converted to plain text using programmatic PDF parsing (no optical character recognition (OCR)), producing minimally processed UTF-8 text files in dataset/bronze/text. Formatting irregularities are preserved at this stage to avoid premature information loss.

3.2.2. Silver Layer: Filename Normalization and Deterministic Parsing

To impose structural consistency, extracted texts are renamed according to a canonical pattern palmares-[year]-[region].txt using an explicit mapping table that resolves historical spelling variants and administrative changes. Deterministic rule-based parsing components (regular expressions and lookup mappings) then identify candidate entries, school headers, and region metadata while preserving a transparent link to the original bronze artifacts.

3.2.3. Gold Layer: Normalization, Structuring, and Curation

Silver-layer texts are reformatted into a uniform, machine-parseable representation via a custom formatter that standardizes whitespace, merges fragmented lines, and enforces a single-line schema for candidate records. School headers are normalized into canonical blocks, with special-case handling for known year-specific layout deviations (e.g., 2023). From the resulting gold text, we extract two primary structured datasets: individual candidate records (names.csv) and school-level statistics (statistics.csv), each annotated with year, region, filename, and source line number.

3.3. Notation and Feature Augmentation

The corpus-construction procedure is fully deterministic. We operate under the following assumptions: (i) palmarès PDFs are predominantly text-based and can be parsed programmatically without OCR; (ii) each candidate entry carries a reported sex marker (M/F) and a region/province association and (iii) name tokenization is whitespace-based after normalization of spacing.

Let $\mathcal{U}$ denote the set of province–year result pages. Crawling $\mathcal{U}$ yields a PDF set $\mathcal{P}$. For any $p\in \mathcal{P}$, let $pdf2txt\left(p\right)$ denote deterministic PDF-to-text conversion (UTF-8, no OCR). For any extracted full name, the whitespace-tokenized form is

$$N=(w_1,w_2,\dots ,w_k),$$

(5)

where $k=\left|N\right|$ is the token count. We define the name complexity function as

$$C\left(N\right)=\left\{\begin{array}{cc}\mathtt{simple},\hfill & \mathrm{if}\left|N\right|=3,\hfill \\ \mathtt{complex},\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(6)

This is a pragmatic structural heuristic: three-token names are the most common structure in the source, but this pattern is not universal and should not be interpreted as a linguistic rule. The character length is

$$\ell \left(N\right)={\displaystyle \sum _{i=1}^k}\left|w_i\right|+(k-1),$$

(7)

where $|w_i|$ counts characters in token $w_i$ and $(k-1)$ accounts for inter-token spaces.

For simple names ($C\left(N\right)=\mathtt{simple}$, i.e., $N=(w_1,w_2,w_3)$), we define a working token-role decomposition for exploratory analysis:

$$G\left(N\right)=(w_1,w_2)\mathrm{and}S\left(N\right)=w_3,$$

(8)

treating the first two tokens as a probable native-name component and the final token as a probable surname component. This is a convenience heuristic, not a ground-truth annotation, and will be inaccurate for some individuals (e.g., compound surnames, variable ordering). For complex names, the decomposition is left undefined. Finally, each record receives a normalized province label via a deterministic region-to-province mapping. The enriched table is exported as names_featured.csv.

3.4. Quality Assurance and Validation

3.4.1. Ablation Study for Regex Refinement

To iteratively improve robustness, we applied the deterministic parsing rules to the gold-layer text, removed matched spans, and inspected the residual text for systematic failure modes (e.g., year-specific formatting changes, wrapped lines, irregular separators). We updated the regex rules accordingly, repeating the cycle until the residual text was dominated by non-candidate content (headers, page artifacts, and school blocks).

3.4.2. Aggregate Extraction Error Rate

We provide a quantitative validation signal by comparing extracted record counts against school-level census figures embedded in the same source PDFs. For each year y, let $E_y$ denote the expected number of passing candidates (summed from school-statistics tables), and let $X_y$ be the extracted passing-candidate row count. The signed extraction error rate is

$$Err\left(y\right)={\displaystyle \frac{E_y-X_y}{E_y}}\times 100.$$

(9)

Positive values indicate under-extraction; negative values indicate over-extraction.

Table 10. Extraction performance and error rate by year.

Year	Entries	Pass	Extracted	Error (%)
2008	316,849	184,887	183,076	0.98
2009	398,896	256,795	256,212	0.23
2010	425,350	282,693	281,536	0.41
2011	472,097	324,245	323,906	0.10
2012	529,949	325,288	324,890	0.12
2013	564,970	271,092	270,705	0.14
2014	611,657	336,206	330,063	1.83
2015	1,207,886	730,338	729,829	0.07
2016	620,671	400,130	400,139	$-0.00$
2017	627,899	412,443	413,563	$-0.27$
2018	677,820	502,045	501,969	0.02
2019	728,688	516,240	516,036	0.04
2020	834,606	633,516	633,454	0.01
2021	1,626,761	956,537	956,994	$-0.05$
2022	1,742,738	1,201,051	1,222,198	$-1.76$
2023	987,341	769,484	709,413	7.81

reports $E_y$, $X_y$, and $Err\left(y\right)$ for all years. Error rates remain below 2% for all years except 2023 (7.81%), which is flagged as lower-confidence due to layout changes in that year’s PDF format; downstream users may wish to exclude 2023 if strict extraction coverage is required.

This metric does not identify which individual rows are erroneous and may be affected by inconsistencies in the published school-statistics blocks. Nevertheless, it provides a concrete, release-level estimate of extraction coverage that is informative for downstream users.

4. User Notes

4.1. Usage Instructions and Recommended Starting Points

The recommended entry point for most use cases is names_featured.csv, which combines candidate records with computed structural features and province labels. For compact name lexicons suitable for NLP tooling, names_unique.csv provides the deduplicated name list. For component-level analysis (e.g., token frequency distributions), names_components.csv provides one row per name token. The statistics.csv file is recommended for studies linking naming patterns to school-level educational outcomes.

All files are stored under dataset/gold/ in the repository. Name strings are normalized to lowercase; applications requiring proper-case output will need to apply re-capitalization heuristics. Complete field-level schemas are provided in Table A1,

Table A2. Schemas of the school-level statistics and deduplicated name datasets. (a) statistics. (b) names_unique.

(a)
Column	Type	Description
index	integer	Row index within source file.
name	string	School name as recorded in source text.
code	string	School code (normalized numeric format).
entries	integer	Total candidates listed.
pass	integer	Candidates who passed.
fail	integer	Candidates who failed.
entries_f	integer	Female candidates listed.
entries_m	integer	Male candidates listed.
pass_f	integer	Female candidates who passed.
pass_m	integer	Male candidates who passed.
fail_f	integer	Female candidates who failed.
fail_m	integer	Male candidates who failed.
year	string	Examination year parsed from filename.
region	string	Region label parsed from filename.
filename	string	Source gold-layer text filename.
(b)
Column	Type	Description
id	string	Identifier of first occurrence.
name	string	Unique full name (deduplicated).
sex	string	Reported sex marker (m or f).
year	string	Examination year parsed from filename.
region	string	Region label parsed from filename.
filename	string	Source gold-layer text filename.
line	integer	Line number of first occurrence.

and

Table A3. Schemas of the derived component-level and unstructured candidate datasets. (a) names_components. (b) names_unstructured.

(a)
Column	Type	Description
id	string	Candidate identifier of original record.
name	string	Full name associated with this component.
sex	string	Reported sex marker (m or f).
year	string	Examination year parsed from filename.
region	string	Region label parsed from filename.
filename	string	Source gold-layer text filename.
line	integer	Line number in gold-layer text file.
full_name	string	Original full name string.
component	string	Single extracted name token.
(b)
Column	Type	Description
name	string	Unmatched candidate-line fragment.
filename	string	Source ablation text filename.
line	integer	Line number in ablation text file.

in the Appendix A.

4.2. Applications and Use Cases

CongoNames supports a broad range of research and applied use cases.

(1)

African NLP and Named Entity Recognition: Large-scale name lexicons derived from this corpus can improve NER systems for French-language and local Congolese language contexts, where coverage of African personal names is typically sparse in standard gazetteers [1]. The corpus can also support spell-checkers, input method editors, and language-model training for Congolese users [15,16].

(2)

Reported Sex Marker Analysis: Because each record carries an administrative sex marker (M/F), the corpus enables analyses of name patterns conditional on that marker. Such studies should frame the sex marker as a proxy derived from administrative records, acknowledging potential inaccuracies. We do not recommend individual-level sex-marker prediction, profiling, or decision-making from names; analyses using this field should be aggregate, hypothesis-driven, and explicit about the distinction between an administrative sex marker and gender identity.

(3)

Regional and Linguistic Variation: Province and year metadata provide an empirical basis for studying regional naming conventions and their correlation with known linguistic geographies, without requiring explicit language-of-origin labels. In future work, we plan to explore adding coarse, explicitly optional “likely language-of-origin” attributions for some name components, via expert-curated lexicons or algorithmic inference.

(4)

Onomastics and Anthropology: The temporal depth (2008–2023) enables diachronic analysis of naming trends, including the effects of the 2015 provincial reorganization on provincial identity and naming practices. The character-level distributions can support generative models (e.g., character-level Markov models or neural language models) for synthetic name generation and NLP stress-testing.

(5)

Data Linkage and Record Matching: The comprehensive Congolese name lexicon can improve the recall of entity-matching algorithms in multilingual database integration scenarios.

(6)

Methodological Transfer: The pipeline methodology is transferable to other countries that publish comparable administrative records. By releasing the processing code openly, we encourage the development of analogous corpora for other underrepresented African nations.

4.3. Limitations

Despite its scale and uniqueness, the corpus is subject to limitations related to population coverage, temporal scope, data quality, normalization choices, and the absence of demographic labels, which should be considered when interpreting the results.

(1)

Coverage Bias: The corpus covers only candidates who reached and sat for the national secondary-school examination in the covered years. This skews toward younger people (late teens) who completed formal secondary education, excluding earlier generations and those outside the formal schooling system. Name distributions may, therefore, differ from the general DRC population.

(2)

Temporal Scope: The current release spans 2008–2023. Earlier decades are not represented, as the ministry did not publish digitized records prior to 2008 and earlier records were paper-based.

(3)

Format Variability: The processing pipeline is tailored to the Examen d’État PDF format. Future changes in ministry layout may require pipeline updates; the 2023 layout change is already flagged as a special case.

(4)

No Ethnicity Labels: Province and year are regional provenance markers, not ethnicity labels. Province is a coarse proxy for geographic origin and is not a reliable indicator of ethnicity, particularly in large, cosmopolitan cities such as Kinshasa.

(5)

Normalization Choices: Lowercasing eliminates capitalization information. Applications requiring properly-cased names must apply re-capitalization heuristics. The heuristic $G\left(N\right)$/$S\left(N\right)$ decomposition is imperfect for individuals with compound surnames, variable ordering, or multi-token family names.

(6)

Residual Extraction Errors: Aggregate error rates are below 2% for most years (Table 10), but residual row-level errors remain possible at the margins. We welcome feedback and will address identified issues in subsequent releases.

4.4. Ethical Considerations and Responsible Use

The source palmarès PDFs are publicly accessible government publications. The released tables contain a limited set of fields—primarily names, a reported sex marker, and year/region provenance—and do not include direct identifiers such as student ID numbers, addresses, phone numbers, or dates of birth. Nevertheless, aggregating names at scale can raise privacy and re-identification risks, particularly because examination candidates are typically minors. We, therefore, release CongoNames for research and educational use only. We encourage aggregate analysis rather than attempts to identify individuals, and we strongly discourage any discriminatory, profiling, or automated decision-making use. Users should comply with applicable laws, institutional review requirements, and data-protection guidelines when combining this resource with other datasets. Public availability of the source examination lists should not be interpreted as removing ethical obligations. The names originate from national exam result publications, include an administrative sex marker, and concern cohorts that largely include underage candidates. Users should avoid reconstructing individual exam histories, linking records to external identity sources, publishing rare-name lookups, or using the corpus for gender inference, surveillance, eligibility screening, or any other individual-level consequential use. When reporting analyses, we recommend aggregation thresholds, suppression of very small cells, and clear statements that the resource is intended for linguistic, NLP, and educational research rather than for identifying or evaluating persons.