saturn

This is the PHOIBLE phonological inventory dataset: 105,484 rows describing phoneme segments across roughly 2,716 language names and 2,177 Glottocodes, with each row carrying a Phoneme/GlyphID plus 40+ binary phonological features (e.g. consonantal, nasal, sonorant, dorsal). The dataset is dominated by consonants — SegmentClass shows 72,282 consonants vs 31,052 vowels and 2,150 tones — and pulls from 8 sources, with 'ph' (36,274) and 'ea' (16,883) accounting for over half the rows. Most feature columns are heavily imbalanced toward '-' or '0', but a handful (consonantal, sonorant, continuant, dorsal, high, front, back) are fairly balanced and carry the real phonological signal worth exploring. Top language names like Iron Ossetic (444), Dutch (395), and Chechen (309) point to the densest inventories in the corpus.

InventoryID looks like a categorical inventory key stored as an integer, with 3,020 distinct values spread across 105,484 rows and no nulls. The distribution is essentially uniform from 1 to 3,020 (mean 1479, median 1464, skew ≈0, kurtosis ≈-1.15), confirming it's an enumerated identifier rather than a measurement. Each ID recurs roughly 35 times on average, so this is a foreign key linking transactions to an inventory dimension.

This column holds Glottocodes — standardized 8-character language identifiers from Glottolog (e.g., 'kham1282', 'dutc1256'). Values are uniformly one word with length tightly clustered at 8 (mean 7.999, min 2, max 8), and there are 2,177 unique codes across 105,484 rows with a 97.9% duplicate rate, indicating each language appears many times. The top code 'kham1282' alone accounts for 622 rows.

This column holds ISO 639-3 language codes: every value is exactly 3 characters and one word, with 2,095 distinct codes across 105,484 rows. The distribution is heavy-tailed and highly repetitive (98.0% duplicate rate), led by 'mis' (828), 'khg' (622), and 'oss' (525), with familiar codes like 'eng' and 'hin' also prominent. No nulls or empties, and the vocabulary (2,086) closely matches n_unique (2,095), consistent with a clean controlled vocabulary.

This column holds language names, mostly single-word labels (one_word_rate 0.84, word_mean 1.20) with a vocab of 2670 across 2716 unique values over 105484 rows. It is highly repetitive (duplicate_rate 0.974) with top entries like 'Iron Ossetic' (444), 'Dutch' (395), and 'Chechen' (309), and roughly 13% of values are uppercase, suggesting inconsistent casing worth normalising. Compound names use directional modifiers ('northern', 'southern', 'eastern', 'western') indicating dialect-level granularity.

Categorical field naming a specific dialect/sub-variety of a language, with 546 distinct values across 105,484 rows. The distribution is extremely concentrated: 71.9% are the literal string "NA" and another 7,692 rows are empty strings, leaving the remaining 544 dialect labels (e.g., "W2", "Lezgian (Güne)", "Scottish Gaelic (Lewis)") in a long tail topping out at 120 occurrences. Entropy ratio of 0.33 confirms most signal lives in that small tail.

GlyphID holds single-token uppercase hexadecimal codes (allcaps_rate 1.0, one_word_rate 1.0, len_min 4, len_median 4) that look like Unicode codepoints — the top values 006D, 0069, 006B map to lowercase Latin letters m, i, k. Despite the ID-sounding name it is highly non-unique: only 3,142 distinct values across 105,484 rows with a 0.97 duplicate rate, so it behaves as a categorical glyph reference rather than a row key. Lengths stretch up to 54 characters (p95 14), hinting that some entries concatenate multiple codepoints.

This column holds individual phoneme tokens, almost always a single character (len_mean 1.50, len_median 1, max 11) and always one word (one_word_rate 1.0). With 105,484 rows but only 3,142 unique values and a 97.0% duplicate rate, the same small phoneme inventory repeats heavily; top symbols like 'm' (2,915), 'i' (2,779), and 'k' (2,729) dominate. Vocab_size of 1,339 is larger than n_unique would suggest for single-character entries, hinting that the longer (up to 11-char) values contribute multi-token strings.

Column holds short phonetic tokens (allophones), with mean length 2.08 characters and 91.3% being a single word. The distribution is dominated by the literal string 'NA' (53,580 of 105,484 rows, ~50.8%), which likely encodes missing rather than true null since null_rate is 0.0; duplicate_rate is 0.935 across only 6,892 uniques. Top non-NA values are individual phoneme letters like 'm', 'j', 'w', 's', consistent with IPA-style symbols.

A ternary flag with values FALSE, NA, and TRUE across 105,484 rows. FALSE dominates at 78.9% while TRUE appears only 1,347 times; the 20,874 NA entries are encoded as a literal string rather than null, so null_rate is 0.0 despite roughly a fifth of rows being missing in practice.

SegmentClass is a categorical phonological label with only 3 distinct values: consonant, vowel, and tone. The distribution is heavily skewed — consonant accounts for 68.5% of 105,484 rows, vowel for 31,052, and tone for just 2,150, giving an entropy ratio of 0.64. The presence of 'tone' as a rare third class suggests the dataset spans tonal languages but those segments are sparsely represented.

Categorical provenance tag with 8 distinct sources across 105,484 rows and no nulls. Distribution is fairly balanced (entropy ratio 0.90), though 'ph' leads at 34.4% followed by 'ea' (16,883) and 'upsid' (13,966). Looks like a dataset-origin code identifying which linguistic database each record came from.

Binary categorical flag with values "0" and "+", almost certainly a tone/sentiment indicator. The distribution is severely imbalanced: "0" covers 97.96% of 105,484 rows while "+" appears only 2,150 times, yielding entropy of just 0.144. No nulls are present.

Binary categorical flag with only two observed values, '-' and '0', across 105484 rows and no nulls. The column is severely imbalanced: '-' covers 97.96% of records (103334) while '0' accounts for the remaining 2150, yielding an entropy ratio of just 0.144. The '-' likely encodes a missing or default state rather than a true category, making this near-constant.

This is a phonological feature column encoding the [syllabic] distinctive feature, with 8 distinct values across 105,484 rows and no nulls. Most entries are simple '-' (68.5%) or '+', but 2,532 rows carry composite codes like '0', '+,-', or '-,+,-' that suggest contour/multi-segment annotations. Entropy ratio of 0.35 confirms the distribution is heavily concentrated on the negative value.

A 4-level categorical with values '-', '0', '+', and '-,+' — likely a strand or sign indicator. It is severely imbalanced: '-' covers 97.76% of 105,484 rows, leaving only 2,150 zeros, 204 plus signs, and 5 mixed '-,+' entries. Entropy ratio is just 0.082, so the column carries almost no information as-is.

Categorical flag with only 6 distinct values dominated by '-' at 89.9% (94844/105484), followed by '+' (8386) and '0' (2150). The remaining three categories are comma-joined combinations like '-,+' and '+,-' with tiny counts (63, 40, 1), suggesting concatenated multi-record values rather than clean single labels. Low entropy ratio (0.214) confirms the column is highly imbalanced toward '-'.

This is a phonological feature column flagging segments as consonantal, with five distinct values across 105,484 rows and no nulls. The vast majority are binary +/- (64,257 and 39,041 respectively, with + dominating at 60.9%), plus 2,151 zero/underspecified entries and just 35 rows with mixed values like '+,-' or '-,+'. Entropy ratio of 0.47 reflects the heavy + skew, and the lone '-,+' singleton is worth noting as a likely encoding artifact.

This is a phonological feature column encoding the [sonorant] value of a segment, dominated by binary '+'/'-' marks (53.0% '+', plus 45322 '-'). The presence of '0' (2150) and comma-joined sequences like '+,-' (1948) and rarer '+,-,-', '+,-,+', '+,-,+,-' suggests contour/complex segments where multiple values are concatenated. Cardinality is 8 with entropy ratio 0.41, so the long tail is sparse but meaningful.

Categorical flag with 9 distinct values across 105,484 rows and no nulls, dominated by '+' (54.9%) and '-' (~42%), with '0' a distant third at 2,151 occurrences. The remaining six categories are comma-joined sequences like '-,+' or '0,0,-,+' that look like concatenated multi-step states rather than atomic labels — only 796 rows total fall into these compound buckets. Entropy ratio of 0.37 confirms heavy concentration in the two primary signs.

A categorical flag named delayedRelease with 7 distinct values across 105484 rows and no nulls. Values are dominated by '0' (55.0%), followed by '-' (27384) and '+' (19533), but a long tail of comma-joined combinations like '-,+', '0,-,+', '+,-', and '0,0,-,+' suggests rows where multiple labels were concatenated rather than normalized. Entropy ratio of 0.52 confirms a skewed but not single-valued distribution.

A low-cardinality categorical with only 6 distinct values dominated by sign tokens '-' (55.9%) and '+' (≈42%), plus a small '0' bucket (2,150 rows) and rare comma-joined combinations like '-,+', '-,-,+', '+,-'. The compound values suggest this field occasionally stores multiple approximant signs concatenated rather than a single label. No nulls across 105,484 rows, and entropy ratio is 0.43, reflecting the heavy '-'/'+' imbalance.

A categorical flag with only 5 distinct values dominated by '-' at 96.7% of 105484 rows, with '0' and '+' as minor categories and two rare composite codes ('-,+', '-,-,+') appearing 25 and 15 times. Entropy ratio of 0.104 confirms extreme imbalance, so this column carries very little discriminative signal on its own. The composite values suggest the field occasionally concatenates multiple states, which is worth verifying against the source schema.

A categorical flag with only 6 distinct values across 105,484 rows, almost certainly encoding a trill or sign indicator ("-", "0", "+", and a few comma-joined sequences). The distribution is severely imbalanced: "-" alone covers 96.15% of rows, leaving entropy at just 0.276 (entropy ratio 0.107). The compound values like "-,+", "-,-,+" and "+,-" appear fewer than 30 times each, hinting at concatenated multi-event records that may need parsing.

A categorical flag for nasal presence/absence, dominated by '-' at 80.8% with '+' a distant second (15,941 of 105,484). The presence of compound values like '+,-', '-,+', and '+,-,-' suggests concatenated multi-observation records rather than a clean binary indicator, and a '0' category (2,150) is a third encoding that doesn't match the +/- scheme.

Categorical flag with 8 distinct values dominated by '-' at 93.8% of 105484 rows, followed by '+' (4211) and '0' (2150). The remaining categories are comma-joined composites like '-,+' or '-,-,+' with counts in the single or double digits, suggesting multi-observation concatenations rather than clean labels. Entropy ratio is just 0.134, so the column carries little information as-is.

A categorical flag for labial articulation, dominated by '-' (68.2%) and '+' (27%), with '0' and various comma-joined combinations like '-,+' and '-,-,+' suggesting multi-segment or sequence-level annotations rather than atomic values. Cardinality is 15 across 105,484 rows with no nulls, and entropy ratio is only 0.30, so the signal is highly concentrated in the binary +/- distinction. The presence of compound tokens is the main surprise and indicates inconsistent encoding granularity.

Categorical column with 8 distinct values dominated by '0' (70.3% of 105484 rows), followed by '+' (16956) and '-' (14082). The tokens look like rounding/sign indicators, and a long tail of compound values like '-,+', '-,-,+', and '0,-,+' (each with 1-269 occurrences) suggests concatenated multi-step rounding sequences. No nulls, but entropy_ratio of 0.398 confirms heavy concentration on the single dominant category.

This column appears to be a phonological feature flag indicating whether a phoneme is labiodental, encoded with values like "0", "-", "+", and a few comma-separated combinations. The distribution is dominated by "0" at 70.3% of 105,484 rows, with "-" at 28,726 and "+" at only 2,574; mixed values like "+,-", "-,+", and "+,+,-" appear in fewer than 60 rows combined and likely represent multi-segment entries. Entropy of 1.01 (ratio 0.39) confirms heavy concentration on the default "0" code.

A low-cardinality categorical with 7 distinct values dominated by sign tokens '-' (62.8%) and '+' (about 36,955 of 105,484), plus a small '0' bucket of 2,160 rows. The presence of comma-joined combinations like '+,-', '-,+', '-,-,+' and '+,-,+' suggests multi-valued entries collapsed into a single string rather than a clean atomic category. Entropy ratio of 0.385 confirms heavy concentration on the two main signs.

A low-cardinality categorical with 6 distinct values dominated by '0' (64.8% of 105484 rows), '+' (25704), and '-' (11391). The remaining three categories are concatenated combinations like '-,+', '+,-', and '-,-,+' totaling only 17 rows, suggesting this field originally allowed multi-valued entries that were collapsed into comma-joined strings. Entropy ratio of 0.48 confirms heavy concentration on the single mode.

A low-cardinality categorical with 11 distinct values dominated by '0' (66.0% of 105,484 rows), followed by '-' and '+'. The remaining eight categories are concatenated combinations like '-,+' or '-,-,+' that together account for fewer than 350 rows, suggesting multi-event encoding squeezed into a single field. Entropy ratio of 0.368 confirms the heavy concentration on the top class.

This appears to be a phonological feature column encoding the [strident] distinctive feature, with values '0' (unspecified), '-' (non-strident), and '+' (strident) covering the vast majority of 105484 rows. About 64.9% are '0' and there are no nulls, but a long tail of 6 composite values like '-,+' and '0,0,-,+' (totaling 625 rows) suggests multi-segment entries where the feature varies across positions. Entropy ratio of 0.41 confirms heavy concentration on the unspecified category.

This column appears to encode a dorsal sign or polarity flag, dominated by two values: '+' (54,535, 51.7%) and '-' (47,052). A third value '0' appears 2,160 times, and the remaining 10 categories are compound comma-separated combinations like '-,+' or '+,-,+', suggesting concatenated multi-observation records collapsed into one cell. Cardinality is 13 with entropy ratio 0.33, so the long tail is negligible but structurally inconsistent with a clean categorical.

This appears to be a categorical 'high' indicator encoding directional movement signs, with '0' (no change) the dominant value at 46.7% of 105,484 rows, followed by '+' and '-'. Compound values like '-,+' and '+,-' suggest concatenated multi-step sign sequences, but they tail off sharply (845, 627, then single digits). Entropy ratio of 0.46 confirms heavy concentration in the top three single-symbol categories; cardinality is 11 with no nulls.

A low-cardinality categorical with 8 distinct tokens dominated by '-' (47.3%) and '0' (~49,244 of 105,484), with '+' a distant third at 5,598. The remaining values are comma-joined sequences like '+,-', '-,+,-', suggesting this column encodes a sign or direction trace, possibly concatenated across events. The long tail (down to a single '+,-,-') indicates rare composite states worth bucketing.

Categorical column encoding a front-side signal with 13 distinct values dominated by three primitives: "0" (49,316), "-" (34,225), and "+" (20,683), together covering nearly all 105,484 rows. The remaining categories are comma-joined combinations like "-,+" (838) or "-,-,+" (24), suggesting multi-event concatenations rather than clean atomic labels. Top rate is 0.468 and entropy ratio 0.43, so the distribution is skewed toward "0" but not degenerate. No nulls.

A low-cardinality categorical with 12 distinct values dominated by the tokens "0" (46.7%), "-", and "+", suggesting a sign/state flag for some "back" attribute. The remaining values are comma-separated combinations like "+,-" or "+,-,-", indicating multiple events concatenated into one cell — a compound encoding rather than a clean atomic category. No nulls across 105,484 rows, and entropy ratio of 0.42 confirms the heavy skew toward "0".

This is a categorical 'tense' field with only 8 distinct values across 105,484 rows and no nulls, dominated by '0' at 71.3% and '+' at ~22%. The remaining categories are sparse markers ('-', '+,-', '-,+') and a long tail of compound sequences with as few as 1-6 occurrences, suggesting a linguistic encoding of tense polarity rather than free text. Low entropy ratio (0.37) confirms heavy concentration in the top class.

Categorical encoding of the retracted-tongue-root (RTR) phonological feature, with 7 distinct values across 105,484 rows and no nulls. The column is severely imbalanced: '-' covers 97.4% of rows, '0' another ~2%, and the remaining five values (including compound codes like '-,+' and '-,-,+') together account for under 500 rows. Entropy ratio of 0.069 confirms almost no information content as-is.

This column encodes the advanced tongue root (ATR) phonological feature, taking three values: '-', '0', and '+'. It is severely imbalanced — '-' covers 97.87% of 105,484 rows, '0' appears 2,235 times, and '+' only 11 times, yielding an entropy ratio of just 0.094. The near-absence of '+' values means this feature carries almost no discriminative signal as-is.

Phonological feature flag for periodic glottal source (voicing), with 7 distinct values across 105,484 rows and no nulls. The vast majority are simple binary tags: '+' at 67.97% (71,694) and '-' (31,179), with a small '0' class (2,139) and rare comma-joined sequences like '+,-' (371) suggesting multi-segment or contour entries. Low entropy ratio (0.3745) confirms heavy concentration on '+'.

A categorical phonological feature (epilaryngeal source) with only 3 distinct values: '-', '0', and '+'. The column is severely imbalanced — '-' accounts for 97.93% of the 105,484 rows, '0' for 2,150 rows, and '+' for just 31 rows, yielding a very low entropy ratio of 0.093. With no nulls but near-constant values, it carries little discriminative signal.

This appears to be a phonological feature column encoding the [spread glottis] distinctive feature, with values '-', '+', '0' and comma-separated combinations for segments with multiple specifications. The distribution is extremely lopsided: '-' covers 91.8% of 105,484 rows and entropy ratio is just 0.149, meaning the column carries little information on its own. The long tail of compound values like '-,+', '+,0,-', and '+,-,+' (some with only 1-5 occurrences) suggests multi-segment or contour entries that may need parsing.

Categorical flag for a 'constricted glottis' phonological feature, with 7 distinct values across 105,484 rows and no nulls. Heavily dominated by '-' at 94.5% (99,727 rows), with '+' and '0' as minor categories and a long tail of comma-joined sequences ('+,-', '-,+', and two singletons) suggesting multi-segment annotations. Low entropy ratio (0.13) confirms the column carries little information in isolation.

A 3-level categorical flag dominated by '-' (68.1% of 105,484 rows), with '0' covering most of the rest and '+' appearing only 415 times. The skew toward '-' and the tiny '+' class (entropy ratio 0.589) suggest a sign/direction indicator rather than a balanced category. No nulls, so the encoding is complete as-is.

A ternary categorical flag with values '-', '0', and '+', likely encoding a linguistic lenition feature (lenis/fortis/neutral). The distribution is highly imbalanced: '-' covers 68.1% of 105,484 rows and '0' another 33,202, while '+' appears only 416 times. No nulls, but the rare '+' class may be too sparse for stable modelling.

This is a categorical phonological feature flag for 'raisedLarynxEjective', taking values like '-', '+', '0', and a few comma-separated combinations across 105484 rows with no nulls. The distribution is severely imbalanced: '-' covers 96.37% of rows and entropy ratio is just 0.103, with rare compound values like '-,-,+' appearing only once. The 6-way cardinality plus mixed-delimiter codes ('-,+' vs '+,-') suggests multi-segment annotations that may need parsing.

Categorical phonological feature flagging lowered larynx implosives, with 5 distinct values across 105,484 rows and no nulls. The distribution is severely imbalanced: '-' covers 97.27% of rows, while '0' (2,150), '+' (716), and the mixed codes '-,+' (7) and '+,-' (2) are rare. Entropy ratio of 0.088 confirms the column carries very little information as-is.

Categorical flag with only 5 distinct values dominated by '-' (68.2% of 105,484 rows) and '0' (33,202), suggesting a click/interaction indicator where '-' likely means no click and '0' a recorded null/zero. The '+' class is rare (253) and the compound values '+,-' (52) and '-,+' (6) hint at concatenated multi-event records that break the single-label assumption. Entropy ratio of 0.40 confirms the heavy imbalance.