joshua project joshua project languages

Overview

Source: /home/coolhand/html/datavis/data_trove/joshua-project/joshua_project_languages.json

Saturn profiled 7,134 rows across 26 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

ROL3 text identifier

ROL3 is a text column of exactly 7134 unique three-character single-word tokens across 7134 rows, with zero nulls and zero duplicates. The perfect 1:1 cardinality (vocab_size 7134 == n) and uniform len_min/max/mean of 3 strongly suggest this is a row-level identifier or code rather than natural language. Top tokens like 'aou', 'aiw', 'aas' show no repeated values, confirming it carries no distributional signal on its own.

Treatment: Drop from modelling or use only as a join key; near-unique three-letter codes carry no predictive signal.

Language text identifier

This column holds language names, with 7124 distinct values across 7134 rows and only 10 duplicates — essentially one row per language. Entries are short (mean 9.1 chars, median 1 word) and 73.5% are single-word labels, though compound names involving directional qualifiers (southern, northern, eastern, western) and family roots (zapotec, mixtec, naga) appear often. The high cardinality combined with the 'language' and 'sign' top tokens suggests this is a catalog of world languages, likely including sign-language variants.

Treatment: Treat as a near-unique label key; left-join on it rather than one-hot encoding.

WebLangText text identifier

WebLangText appears to be a per-row language name label, with every one of the 7134 values unique and 73% being a single word (mean 1.37 words, median length 7 chars). Top tokens like 'language', 'sign', 'zapotec', 'mixtec', and 'naga' suggest this is an inventory of world languages including sign languages and regional variants (Southern/Northern/Eastern/Western). The full uniqueness (n_unique == n) means it functions as an identifier rather than a categorical feature.

Treatment: Treat as a language-name key; left-join on this rather than using as a model feature.

Status categorical label

Binary status flag with two values, 'L' and 'N', dominated by 'L' at 86.0% (6134 of 7134) versus 'N' at 998. Class imbalance is notable, and there are 2 nulls (null_rate 0.0003). Entropy ratio of 0.58 confirms the skewed distribution.

Treatment: Encode as binary; address class imbalance (e.g., stratified sampling or class weights) before modelling.

ROG3 categorical feature

ROG3 is a categorical code field with 211 distinct two-letter values, dominated by 'PP' at 11.7% (837 rows) followed by 'ID', 'NI', 'IN', 'MX' — a distribution consistent with country or region codes. Entropy is 5.64 (ratio 0.73), indicating broad spread across the 211 categories rather than concentration in a few. Nulls are negligible (0.03%) and the top-10 values mix what look like ISO-style country codes ('IN', 'MX', 'US', 'CH') with less standard tokens ('PP', 'NI', 'CG').

Treatment: Target-encode or group rare levels before modelling; verify whether codes follow an ISO standard.

HubCountry categorical feature

HubCountry is a categorical country-name field with 210 distinct values across 7,134 rows and a near-zero null rate (0.0003). The distribution is broad rather than concentrated — entropy ratio 0.73, with the top value Papua New Guinea covering only 11.7% of records, followed by Indonesia (686) and Nigeria (494). The leading countries skew toward biodiversity- or resource-rich nations rather than the largest economies, which is worth noting before any geographic modelling.

Treatment: Group long-tail countries into regions or frequency buckets before one-hot or target encoding.

BibleStatus numeric feature

BibleStatus is an integer-coded categorical with 6 distinct values from 0 to 5, mean 2.68 and median 3, almost certainly an ordinal status/level code rather than a true numeric measure. About 14.9% of rows are zero and the distribution is mildly left-skewed (skew -0.34) with flat kurtosis (-0.91), suggesting a fairly even spread across the upper levels with a sizable zero/'none' bucket. Null rate is negligible (0.0003) and no outliers were flagged.

Treatment: Treat as an ordinal categorical (one-hot or ordered encoding) rather than a continuous numeric.

GRN_URL text identifier

This column holds Global Recordings Network language URLs, every value a fixed 44-character single token under https://globalrecordings.net/en/language/ followed by a language code. With 4179 unique values across 7134 rows and a 41.41% null rate, it functions as a per-language identifier link rather than a feature. Notable: only one duplicate URL (idt appears twice) despite the high uniqueness, and 41% of rows have no GRN link at all.

Treatment: Extract the trailing language code as a foreign key; otherwise drop the URL itself from modelling.

TranslationNeedQuestionable unknown other

The column 'TranslationNeedQuestionable' was skipped by the profiler, so no type, cardinality, or value statistics are available beyond a row count of 7134 and a null rate of 0.0. The name suggests a boolean or flag indicating whether the need for translation is in doubt, but this cannot be confirmed from the evidence. No distribution, unique count, or sample values were captured.

Treatment: Re-profile with type inference enabled before deciding on downstream use.

BibleYear categorical free_text

Free-text field nominally capturing a year associated with a Bible (likely year acquired or published), but heavily polluted: 89.01% of 7134 rows are null, and among the 488 unique values the most common entry is "2023" at just 3.4%, with "Yes" appearing as the second most frequent value (22 times) — indicating the field also absorbed yes/no answers. Entropy ratio of 0.93 confirms a long, flat tail with no dominant year.

Treatment: Clean by coercing to integer years and routing non-numeric responses (e.g., "Yes") to a separate flag before use.

NTYear text feature

Despite the name suggesting a year, NTYear is a single-token text column with mixed semantics: the most frequent value is 'Yes' (147 rows), followed by four-digit years from 2016-2024. It's 63.61% null and 57.28% duplicates across only 1109 unique values, with one_word_rate of 1.0 and allcaps_rate of 0.94. The mix of a yes/no token alongside year values suggests two questions were collapsed into one field.

Treatment: Split into two columns (boolean indicator vs. numeric year) before any modelling.

PortionsYear text feature

Despite the name PortionsYear, this column mixes a yes/no flag with four-digit years: every entry is a single word, the most common value is 'Yes' (706 occurrences) followed by years like 2024 (107), 2022 (55) and 2025 (43). 43% of rows are null and 55.8% of the non-null values are duplicates across only 1,797 unique tokens, with 82.6% in all-caps. The semantic mix of a boolean and a year in one field is the headline anomaly.

Treatment: Split into two columns — a boolean 'has portions' flag and a parsed year — before modelling.

PercentAdherents numeric feature

PercentAdherents is a numeric share variable bounded between 0 and 100, almost certainly the percentage of some population that adheres to a religion or group. The distribution is strongly bimodal in feel: the IQR spans 5.34 to 90.0 with a median of 58.33, kurtosis of -1.68 indicates a flat/U-shaped spread rather than a central peak, and 7.3% of values are exactly zero. Nearly 12% of rows are null, which is worth flagging before any aggregation.

Treatment: Impute or filter the 11.87% nulls and consider binning given the U-shaped distribution before modelling.

PercentEvangelical numeric feature

This appears to be a percentage feature capturing the share of evangelicals in each record's population, ranging from 0 to 95 with a median of just 5. The distribution is heavily right-skewed (skew 1.93, kurtosis 4.92) and 9% of values are exact zeros, with 251 outliers (4.3%) pulling the mean up to 10.1. Note also that 17.3% of rows are null, which is substantial.

Treatment: Impute or flag the 17% nulls and consider a log1p transform before modelling to tame the right skew.

HasJesusFilm categorical feature

Binary Y/N flag indicating whether each record has an associated Jesus Film, with only 2 unique values across 7134 rows and a negligible 0.0003 null rate. The distribution is moderately imbalanced: 'N' dominates at 71.6% (5105) versus 2027 'Y' values, yielding an entropy ratio of 0.86.

Treatment: Encode as a 0/1 boolean indicator for modelling.

JF_URL text metadata

This column holds JesusFilm.org URLs (url_rate 1.0, one_word_rate 0.9995), almost all pointing to language-specific watch pages like /watch/jesus.html/{language}.html. Coverage is sparse with a 71.6% null rate, and values are near-unique (2008 distinct out of 7134), though a generic partners/resources page appears 13 times and 19 duplicates exist overall. URL lengths are tight (45-87 chars, median 55), consistent with a templated link rather than free text.

Treatment: Treat as a reference link; extract the language slug from the path if you need a feature, otherwise drop from modelling.

HasAudioRecordings categorical feature

Binary Y/N flag indicating whether a record has associated audio recordings. The split is 59.1% Y vs N, with entropy ratio 0.976 showing near-maximal balance for a two-class field. Null rate is negligible (0.0003).

Treatment: Encode as a boolean indicator before modelling.

JPScale categorical feature

JPScale is a low-cardinality categorical with 5 distinct values ('1' through '5'), suggesting an ordinal rating or scale. Distribution is bimodal: the extremes '5' (32.6%) and '4' (26.6%) dominate alongside '1' (20.6%), while middle values '3' and '2' are comparatively rare. Entropy ratio of 0.89 indicates fairly even spread across categories, but 11.87% of rows are null.

Treatment: Treat as ordinal (1–5); impute or flag the ~12% nulls before modelling.

LeastReached categorical feature

Binary Y/N flag indicating whether some 'least reached' status applies, with N dominating at 79.3% (5659) versus 1473 Y values across 7134 rows. Class imbalance is notable but not extreme, and nulls are negligible (0.03%). Cardinality is exactly 2 with entropy ratio 0.73, consistent with a clean boolean indicator.

Treatment: Encode as boolean (Y=1, N=0) and account for the ~80/20 imbalance if used as a target.

RLG3 numeric feature

RLG3 is a small-cardinality integer-coded numeric (only 8 unique values across 7134 rows, ranging 1-9 with no zeros), suggesting an ordinal scale or category code rather than a true continuous measure. The distribution leans low: median 1, Q3 of 4, mean 2.82, with right skew (0.74) and 110 outliers (1.7%) toward the high end. Note the 10.88% null rate, which is non-trivial and should be addressed before modelling.

Treatment: Treat as ordinal/categorical and impute or flag the ~11% missing values before modelling.

PrimaryReligion categorical feature

Categorical label for the dominant religion of each record, with 9 distinct values across 7134 rows. Christianity leads at 46.7% (3328), followed by Ethnic Religions (1472) and Islam (945). Note the empty-string category appears 774 times alongside an explicit 'Unknown' bucket of 110 — two separate missingness conventions coexist beyond the 0.03% true nulls.

Treatment: Consolidate the empty string into 'Unknown' and one-hot encode the 9 categories.

FCBH_URL text metadata

This column holds a single URL per row pointing to Faith Comes By Hearing resources (apk.fcbh.org or live.bible.is), with url_rate at 1.0 and one_word_rate at 0.9987. It is largely missing — null_rate is 0.6801 — and of the populated rows 2272 of values are unique with only 10 duplicates. Lengths are tight (min 25, max 100, mean 37.68), consistent with a structured link field rather than free text.

Treatment: Treat as an optional reference link; keep as-is for lookup, do not feed into modelling.

NbrPGICs numeric feature

NbrPGICs is a heavily right-skewed count feature, with median 1 and Q3 of 2 but a maximum of 1804 and standard deviation of 48.93. The distribution shows extreme tail behaviour (skew 16.65, kurtosis 404.08) and 701 outliers (11.0% of values), while 10.75% of rows are null. Most records carry a trivial count, but a small subset reports values orders of magnitude larger.

Treatment: Log-transform or cap at a high quantile and impute the 10.75% nulls before modelling.

NbrCountries numeric feature

NbrCountries is a numeric count of countries associated with each record, ranging from 1 to 136 with a median of 1 and Q1=Q3=1, meaning at least three quarters of rows are single-country. The distribution is extremely heavy-tailed (skew 15.6, kurtosis 364) with 1203 outliers (20.5% outlier rate) and a 17.75% null rate, so a small minority of multi-country records dominate the variance.

Treatment: Log1p-transform or bucket into 1 vs. multi-country before modelling, and impute or flag the 17.75% nulls.

JF categorical feature

Binary Y/N flag with only two distinct values across 7134 rows and a negligible null rate of 0.0003. The distribution is skewed toward 'N' at 71.6%, leaving 'Y' at roughly 2027 occurrences. Entropy ratio of 0.86 indicates the split is imbalanced but still informative.

Treatment: Encode as a 0/1 indicator and impute the rare nulls with the mode.

AudioRecordings categorical feature

Binary Y/N flag indicating whether audio recordings exist for each row, with only 2 unique values across 7,134 records and a negligible null rate of 0.0003. The split is moderately balanced toward 'Y' at 59.1% (4,216) versus 'N' (2,916), giving high entropy (0.976) for a binary field. No surprising signals beyond the slight Y-majority skew.

Treatment: Encode as a 0/1 boolean indicator after imputing the few nulls.