data trove phoible phonetics database

Overview

Source: /home/coolhand/html/datavis/data_trove/data/linguistic/phoible.csv

Saturn profiled 105,484 rows across 49 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).

InventoryID numeric foreign_key

InventoryID is a numeric foreign key referencing an inventory dimension table, with exactly 3,020 distinct values spanning 1–3,020 across 105,484 rows, implying heavy repeated use of each ID (average ~35 rows per ID). The distribution is remarkably flat and symmetric (skew ≈ −0.002, kurtosis ≈ −1.15, zero outliers), consistent with a well-populated lookup identifier rather than a measured quantity.

Treatment: Left-join on this ID to the inventory dimension table; do not use raw numeric value as a feature.

Glottocode text foreign_key

This column contains Glottocodes — the standardized 8-character language identifiers used by the Glottolog database (e.g., 'kham1282', 'dutc1256'), confirmed by the near-uniform length of 8 characters (mean 7.999, median 8.0) and the structured alphanumeric format. With only 2,177 unique codes across 105,484 rows, the duplicate rate is extremely high at 97.9%, meaning each language code recurs on average ~48 times — consistent with a dataset where many observations (e.g., words, features, speakers) are annotated per language. The top code 'kham1282' (Kham) appears 622 times, suggesting uneven language coverage in the dataset.

Treatment: Left-join on this code against the Glottolog reference table to enrich with language family, geographic coordinates, and macro-area metadata.

ISO6393 text label

This column contains ISO 639-3 three-letter language codes, a standardised identifier for natural languages. Every value is exactly 3 characters long (min, mean, max all equal 3) with zero nulls, confirming strict conformity to the standard. The duplicate rate is extremely high at 98.0%, meaning 2,095 distinct codes repeat across 105,484 rows — expected behaviour for a language tag applied to many records. The most frequent code 'mis' (miscellaneous/unattested language) appearing 828 times may warrant attention, as it signals a non-trivial share of records with unidentified languages.

Treatment: Use as a categorical grouping key; consider flagging or separating records where ISO6393 equals 'mis' (unattested) before language-based analysis.

LanguageName text label

This column contains human language names (e.g., 'Dutch', 'Chechen', 'Bengali', 'Iron Ossetic'), functioning as a categorical label drawn from a linguistically diverse vocabulary of 2,716 unique values across 105,484 rows. The duplicate rate of 97.4% (102,768 duplicates) confirms it is a low-cardinality repeating label, not a free-text field. Notably, 13.1% of values are all-caps, suggesting some entries use ISO-style abbreviations or codes alongside full names. The distribution is uneven — the top value 'Iron Ossetic' appears 444 times while many languages appear rarely — indicating a long-tail spread across minority and regional languages.

Treatment: Encode as a categorical (label or target-encode) after normalising casing inconsistencies flagged by the 13.1% all-caps rate.

SpecificDialect categorical label

This column captures a specific dialect designation for each record in what appears to be a linguistic or language survey dataset. The dominant signal is extreme missingness-by-label: 71.9% of rows carry the value 'NA' and a further 7,692 rows (≈7.3%) are empty strings, meaning roughly 79% of records lack a meaningful dialect value. Despite 546 unique values, the effective entropy ratio is only 0.33, confirming that real dialect labels are thinly and unevenly spread across the remaining ~22,000 rows.

Treatment: Treat 'NA' and empty-string as missing; consider collapsing rare dialects (below a frequency threshold) into an 'Other' bucket before encoding, and flag high missingness rate (~79%) before any modelling use.

GlyphID text label

GlyphID is a categorical identifier column containing Unicode code point values (e.g., '006D' = 'm', '0069' = 'i', '0061' = 'a') stored as uppercase hexadecimal strings, almost certainly representing character-level glyph references in a typography, OCR, or text-rendering dataset. With only 3,142 unique values across 105,484 rows, the duplicate rate is extremely high at 97.0%, which is expected for a glyph/character frequency table but means this column functions as a low-cardinality label rather than a unique identifier despite its name. All values are fully uppercase (allcaps_rate 1.0), single-token (one_word_rate 1.0), and short (median length 4 characters), consistent with 4–6 character hex codes. The top values map to common Latin lowercase letters, suggesting the underlying corpus is predominantly Latin-script text.

Treatment: Map hex strings to Unicode characters for interpretability, then encode as categorical (low-cardinality, 3142 levels) or group by Unicode block/script before modelling.

Phoneme text label

This column contains phoneme symbols from a linguistic or speech dataset, with 3,142 unique phoneme strings across 105,484 rows. Values are overwhelmingly single characters (len_median 1.0, len_mean 1.5, len_max 11), consistent with IPA or ARPABET-style phoneme notation. The duplicate rate is 97.0% (102,342 duplicates), which is expected given a finite phoneme inventory repeated across many words or utterances. The vocab_size of 1,339 distinct tokens against only 3,142 unique values suggests multi-character phoneme strings (e.g., digraphs or diacritics) are also present alongside single-character ones.

Treatment: Encode as categorical (label or one-hot) for modelling; 3,142 unique values is manageable but consider grouping by manner/place of articulation if dimensionality is a concern.

Allophones text feature

This column contains allophone representations from a phonology or linguistics dataset, storing the phonetic variants of phonemes (e.g., 'm', 'j', 'w', 's') as very short strings with a mean length of ~2 characters. Strikingly, 53,580 of 105,484 rows (roughly 50.8%) carry the sentinel value 'NA', indicating no allophone recorded, which dominates the duplicate rate of 93.5%. The 6,892 unique values across a vocabulary of only 1,263 words suggest a modest set of phonetic symbols combined in small clusters, consistent with IPA or similar notation.

Treatment: Treat 'NA' as missing; encode remaining values as categorical or tokenize individual phoneme symbols before modelling.

Marginal categorical feature

This column is a boolean flag indicating whether a record is classified as 'marginal', with three distinct string values: FALSE, NA, and TRUE. The dominant value is FALSE (78.9% of 105,484 rows), while TRUE is strikingly rare at only 1,347 records (≈1.3%). The presence of 'NA' as a literal string value in 20,874 rows (≈19.8%) is noteworthy — these are not system nulls (null_rate is 0.0) but encoded string missings, which must be handled explicitly rather than via standard null imputation.

Treatment: Encode as ternary (FALSE=0, TRUE=1, NA=missing) after converting string 'NA' to actual nulls, then decide on imputation strategy given 19.8% string-missing rate.

SegmentClass categorical label

SegmentClass is a phonological category label classifying 105,484 linguistic segments into exactly three classes: consonant, vowel, and tone. Consonants dominate at 68.5% (72,282 occurrences), vowels account for 29.4% (31,052), and tones are a small minority at just 2.0% (2,150) — a distribution consistent with natural language phoneme inventories but with tones notably underrepresented, suggesting the dataset skews toward non-tonal languages or tonal markings are partially absent. Zero nulls and perfect coverage make this a clean, reliable label.

Treatment: One-hot encode or use as a stratification variable; monitor class imbalance for the 'tone' minority class (2,150 samples) in any classification task.

Source categorical label

This column identifies the source database or linguistic inventory from which each phonological record was drawn, with 8 distinct source codes across 105,484 rows and no nulls. The top source 'ph' dominates at 34.4% (36,274 rows), likely referring to PHOIBLE or a similar phoneme database, followed by 'ea', 'upsid', 'er', 'saphon', 'aa', 'spa', and 'ra'. The high entropy ratio of 0.899 indicates a relatively even spread across sources despite 'ph' leading — no single source overwhelmingly controls the data. Analysts should be aware that cross-source comparisons may introduce systematic coding differences, as each source may apply distinct phonological conventions.

Treatment: One-hot encode or use as a stratification/grouping variable; check for source-specific biases before pooling.

tone categorical label

This column is a binary tone indicator with only two values: '0' (neutral/absent) and '+' (positive), across 105,484 rows with no nulls. The distribution is severely imbalanced — '0' accounts for 97.96% of records (103,334) while '+' appears in just 2,150 rows (~2%). The entropy of 0.144 (out of a maximum of 1.0 for a binary variable) confirms the near-total dominance of a single class, which would make any model trained on this label prone to predicting '0' exclusively.

Treatment: Apply class-balancing techniques (e.g., oversampling '+' or class-weight adjustment) before using as a modelling target.

stress categorical label

This column is a binary categorical flag likely indicating the presence or absence of stress (e.g., a linguistic or physiological stress marker), encoded as '-' for the negative/absent case and '0' for the positive/present case. The distribution is severely imbalanced: '-' accounts for 97.96% of all 105,484 rows (103,334 occurrences) versus only 2,150 rows for '0'. The entropy ratio of 0.144 confirms near-minimum information content, meaning the positive class is rare and models trained on this column will need class-balancing strategies.

Treatment: Apply class-balancing (oversampling or weighted loss) before modelling; consider encoding '-' as 0 and '0' as 1 for numeric compatibility.

syllabic categorical

short categorical feature

This column appears to encode a directional or sign indicator with only 4 distinct values: '-', '0', '+', and the compound '-,+'. It is severely imbalanced: the dominant value '-' accounts for 97,764% of rows (103,125 of 105,484), while '+' appears just 204 times and the mixed label '-,+' only 5 times. The near-zero entropy ratio (0.082) confirms the column carries very little information variance, and any model trained on it will be overwhelmed by the '-' class.

Treatment: Treat as ordinal or one-hot encode, but flag severe class imbalance (97.8% '-'); consider oversampling minority classes or collapsing '+' and '-,+' before modelling.

long categorical feature

This column encodes the sign or direction of longitude (or a signed numeric field), with only 6 distinct values across 105,484 rows. The dominant value '-' accounts for 89.9% of records, suggesting the dataset is heavily skewed toward negative longitudes (e.g., Western Hemisphere coordinates). The compound values '-,+', '+,-', and '-,-,+' appear in just 104 rows combined, hinting at rare multi-value or malformed entries that warrant inspection before modelling.

Treatment: Investigate compound values ('-,+', '+,-', '-,-,+') for parsing errors; encode '-' as -1, '+' as +1, '0' as 0, and flag or impute the 104 compound rows.

consonantal categorical feature

This column encodes a binary phonological feature — consonantal — marking whether a segment is consonantal (+) or not (-), a standard distinctive feature in linguistics datasets. The dominant value is '+' at 60.9% (64,257 rows), with '-' covering 37.0% (39,041 rows). The near-zero third category '0' (2,151 rows) suggests underspecified segments, while the composite values '+,-' (34) and '-,+' (1) are anomalous multi-value entries that likely reflect data quality issues or ambiguous annotations.

Treatment: Encode as ordinal or one-hot after splitting composite entries ('+,-', '-,+') into separate flags or flagging them for review.

sonorant categorical

continuant categorical feature

This column encodes a phonological feature — whether a speech sound is a continuant (airflow continues through the vocal tract) — using '+', '-', and '0' (unspecified/N-A) notation standard in distinctive feature theory. The dominant values are '+' (54.9%, 57,952 rows) and '-' (44,585 rows), which together account for ~97% of records. Surprisingly, 7 of the 9 unique values are composite strings like '-,+' or '-,-,+', suggesting some rows encode sequences or multi-segment entries rather than single-segment features, which is atypical and may indicate data quality or schema inconsistency. Entropy ratio of 0.37 reflects moderate concentration driven almost entirely by the binary +/- split.

Treatment: Treat '+'/'-'/'0' as a 3-class categorical; investigate and potentially split or flag the 796 composite multi-value rows before encoding.

delayedRelease categorical feature

This column encodes a delayed-release flag or classification for 105,484 records using a small set of symbolic tokens: '0' (no delay, 55% of rows), '-' (negative/early shift, 26%), '+' (positive/late shift, 19%), and compound combinations thereof. The compound values ('-,+', '0,-,+', '+,-', '0,0,-,+') suggest the column is sometimes multi-valued — packed as comma-separated strings — which is structurally inconsistent with a simple categorical and implies set-like semantics. No nulls are present, and the entropy ratio of 0.52 reflects a moderately skewed but non-trivial distribution dominated by the '0' class.

Treatment: Split compound values on ',' into multi-hot binary columns ('has_0', 'has_minus', 'has_plus') before modelling.

approximant categorical feature

This column captures whether a phoneme or linguistic segment is classified as an approximant, using a compact symbolic encoding where '-' means absent and '+' means present. The dominant value is '-' at 55.9% of rows (58,966), with '+' accounting for 41.9% (44,266), giving a near-binary distribution with very low entropy (1.12). Surprising are the small number of compound values ('-,+', '-,-,+', '+,-') totalling just 102 rows, suggesting a minority of segments carry ambiguous or multi-valued approximant classifications — possibly transcription artifacts or multi-segment bundles.

Treatment: One-hot encode the two dominant values ('-', '+'); bin the three compound values into an 'ambiguous' category or flag and investigate as potential data quality issues before modelling.

tap categorical feature

This column appears to encode a directional or signed-change indicator, with values like '-', '+', '0', and compound sequences such as '-,+' and '-,-,+' that suggest a history of sign changes or transitions. The dominant value '-' accounts for 96.7% of all 105,484 rows, producing severe class imbalance (entropy ratio of only 0.10), meaning positive or mixed-direction events are rare signals. The compound multi-value entries ('-,+', '-,-,+') are notable — they imply the field can store ordered sequences of transitions, not just a single state. No nulls are present.

Treatment: One-hot encode single-value categories; treat compound sequence values ('-,+', '-,-,+') separately via sequence parsing or a dedicated binary flag for multi-transition events; oversample or weight minority classes before modelling.

trill categorical feature

This column encodes a phonological feature — specifically the presence or absence of a trill articulation — using a small symbol vocabulary of 6 distinct values. The dominant value is "-" (absent/negative), which accounts for 96.15% of 105,484 rows, producing an extremely low entropy of 0.276 and triggering an imbalance alert. The minority values ("0", "+", and compound sequences like "-,+") together cover fewer than 4,000 rows, suggesting trill is a rare feature in this phonological dataset. The compound values ("-,+", "-,-,+", "+,-") with counts of 26, 8, and 2 respectively hint at multi-segment or allophonic annotation, but are statistically negligible.

Treatment: One-hot or ordinal encode with caution; severe class imbalance (96.15% negative class) means most models will need oversampling, class weighting, or collapsing rare compound categories before use.

nasal categorical

lateral categorical feature

This column encodes lateral direction or laterality coding, likely indicating the side(s) of a clinical finding, anatomical measurement, or test result, using '-' (e.g., left or negative), '+' (right or positive), and '0' (midline or neutral). The dominant value '-' accounts for 93.8% of all 105,484 rows, producing a very low entropy ratio of 0.134, meaning the column is heavily skewed toward a single class. Surprisingly, some values represent sequences of laterality codes (e.g., '-,+', '-,-,+'), suggesting multi-segment or bilateral recordings encoded as comma-delimited strings rather than separate fields. No nulls are present.

Treatment: Split comma-delimited compound values into multi-hot binary flags for '-', '+', and '0' presence before modelling; expect severe class imbalance.

labial categorical feature

This column captures labial articulation features in a phonological or linguistic dataset, encoding presence/absence/neutrality of a labial feature per segment (or per segment sequence). The dominant value is '-' (absent) at 68.2% of 105,484 rows, with '+' (present) at 26.8%, suggesting most segments are non-labial. Surprisingly, ~2.9% of values are compound strings like '-,+', '+,-', '-,-,+', indicating multi-segment bundles packed into a single cell rather than a flat atomic feature, which creates a parsing challenge and implies the column is not fully normalized.

Treatment: Split compound values on ',' into separate per-segment records or encode as ordered multi-label; then one-hot or ordinal encode the atomic {'-', '+', '0'} values for modelling.

round categorical

labiodental categorical feature

This column encodes labiodental phonetic feature annotations, likely from a linguistic or phonological dataset where each row represents a speech sound or segment. The dominant value '0' (74,124 rows, 70.3%) indicates the feature is absent or neutral, while '+' and '-' mark positive and negative feature values respectively — a standard binary distinctive-feature notation. Notably, multi-valued strings like '+,-', '-,+', and '+,+,-' appear (60 rows total), suggesting a small number of segments carry conflicting or composite annotations, which may indicate data entry inconsistency or deliberate underspecification. Entropy ratio of 0.39 confirms moderate imbalance with '0' dominating.

Treatment: One-hot encode '+', '-', '0'; flag or inspect the 60 multi-valued rows ('+,-', '-,+', '+,+,-') for normalization before modelling.

coronal categorical feature

This column encodes coronal articulation features in what appears to be a phonological or linguistic dataset, using binary '+'/'-' notation common in distinctive feature theory. The dominant value is '-' (62.8% of 105,484 rows), with '+' covering another 35.0%, making the two primary values account for ~98% of observations. The remaining values ('+,-', '-,+', '-,-,+', '+,-,+') suggest multi-segment or compound entries, which are rare (≤87 occurrences) but worth flagging as potential encoding inconsistencies or multi-value cells that may need splitting.

Treatment: Encode primary values '+'/'-'/'0' as ordinal or one-hot features; isolate and inspect the 135 multi-value rows for parsing or splitting before modelling.

anterior categorical feature

This column captures an anterior anatomical or directional finding coded as a signed categorical variable, with 6 distinct values across 105,484 rows and zero nulls. The dominant value '0' (64.8% of rows) likely indicates absence or neutral status, while '+' (24.4%) and '-' (10.8%) denote positive/negative findings. Surprisingly, a small number of rows (9, 5, and 3) contain compound multi-value strings like '-,+', '+,-', and '-,-,+', suggesting occasional data-entry concatenation errors or multi-event encoding that deviates from the expected single-symbol schema.

Treatment: Map '0'/'+'/'-' to ordinal or one-hot encoded features; isolate and investigate the 17 compound-value rows before modelling.

distributed categorical label

This column appears to encode a distribution or change-direction indicator, using a compact symbolic notation: '0' (no change), '-' (decrease), '+' (increase), and comma-separated sequences for multi-step or compound movements. The dominant value '0' covers 66% of 105,484 rows, with '-' (21.1%) and '+' (12.5%) accounting for most of the remainder. Surprisingly, 11 distinct values arise from combinations of these three symbols, suggesting some records capture sequences of directional changes rather than a single state. Entropy ratio of 0.37 confirms moderate but uneven information content, heavily concentrated in the single '0' class.

Treatment: Encode as an ordinal or multi-hot feature by splitting on commas and mapping {'-': -1, '0': 0, '+': 1}; treat compound values as sequences.

strident categorical

dorsal categorical feature

This column captures a dorsal-surface marking or pattern indicator for biological specimens (likely fish or reptiles), encoding the presence/absence of a feature using '+', '-', and '0' symbols — sometimes as comma-separated sequences indicating multiple zones or segments. The dominant values are '+' (51.7%, n=54,535) and '-' (n=47,052), together accounting for ~96.3% of rows, with 'neutral/absent' coded as '0' (n=2,160). Surprisingly, 11 of the 13 categories are multi-value strings like '-,+' or '-,-,+', suggesting some records encode ordered sequences of dorsal sub-regions rather than a single binary state, creating an implicit structural inconsistency that warrants normalization.

Treatment: Split multi-value entries on ',' into ordered sub-features or one-hot encode each position before modelling; treat '+', '-', '0' as a ternary categorical for single-value rows.

high categorical feature

This column encodes the direction of price/value movement at a 'high' point — likely a signal or pattern indicator from a time-series or financial dataset, using tokens '0' (no movement), '+' (up), and '-' (down), including multi-step sequences like '-,+' or '+,-,+'. The dominant value is '0' at 46.7% (49,247 rows), followed by '+' at 35,559 and '-' at 19,156, suggesting a strong asymmetry between upward and downward signals. The presence of compound sequence values (e.g., '+,-,+', '-,+,+') with very low frequency — some appearing only once or twice — indicates these multi-step patterns are rare edge cases that may need consolidation or separate treatment.

Treatment: Encode top 3 values ('0', '+', '-') as ordinal or one-hot features; collapse rare multi-step sequences (≤845 occurrences) into an 'other' or structured sub-category.

low categorical

front categorical feature

This column appears to encode a directional or signed-change indicator for a 'front' measurement, using values '0', '-', and '+' as atomic tokens that can be combined into sequences (e.g., '-,+', '+,-,-'). The dominant value '0' accounts for 46.75% of rows, with '-' (34,225) and '+' (20,683) making up most of the remainder — notably, negative changes outnumber positive ones by ~1.65:1. The compound sequence values ('-,+', '+,-', etc.) suggest multi-step event chains are encoded as a single string, which is an unusual encoding pattern that may require parsing before modelling.

Treatment: Parse composite sequences (e.g., '-,+') into structured event arrays or ordinal counts; then one-hot or embed atomic states before modelling.

back categorical feature

This column encodes directional movement or change signals for a 'back' dimension, using a compact notation of '+', '-', and '0' tokens — likely representing price/value movement sequences (e.g., up, down, flat) over a lookback window. The dominant value is '0' at 46.7% (49,270 rows), followed by '-' at 39,749 and '+' at 15,547, indicating a strong bearish/negative skew relative to positive signals. Compound multi-step sequences like '+,-' (511) and '-,+' (367) exist but are rare, and a handful of three-step sequences appear at the tail, suggesting variable-length encoding that could cause parsing issues if treated as a simple label.

Treatment: Parse composite multi-token values (e.g., '+,-') into structured sequences or ordinal scores before modelling; consider one-hot or frequency encoding for the simple single-token majority.

tense categorical

retractedTongueRoot categorical feature

This column encodes the phonetic feature 'retracted tongue root' (RTR), a binary or multi-valued linguistic annotation used in phonological databases to mark vowel or consonant articulation. The dominant value is '-' (absence of RTR) at 97.44% of 105,484 rows, making the feature extremely rare in this dataset — only ~2,696 tokens show any positive marking. The compound values ('-,+', '-,-,+', etc.) suggest per-segment annotation strings rather than single-token labels, indicating a sequence or multi-segment scope.

Treatment: Flag severe class imbalance (97.44% negative); use oversampling or class-weighted models if predicting RTR; consider splitting compound strings into per-segment binary indicators before modelling.

advancedTongueRoot categorical feature

This column encodes the Advanced Tongue Root (ATR) phonological feature, a binary articulatory distinction marked as '+' (present), '-' (absent), or '0' (neutral/not applicable), typical of linguistic phoneme inventories. The distribution is severely imbalanced: '-' dominates at 97.87% (103,238 of 105,484 rows), '0' appears in only 2,235 rows, and '+' is nearly absent with just 11 occurrences. The near-zero entropy ratio (0.094) confirms that '+ATR' is a vanishingly rare feature in this dataset, which would make any model predicting '+' extremely difficult to train without resampling.

Treatment: Treat as ordinal or one-hot encoded categorical; oversample or reweight the '+' class (n=11) before modelling, or collapse '+' and '0' if class separation is infeasible.

periodicGlottalSource categorical

epilaryngealSource categorical feature

This column encodes whether an epilaryngeal phonation source is present, using a ternary scheme: absent ('-'), present ('+')), or a neutral/ambiguous state ('0'). The distribution is severely imbalanced: the '-' (absent) class dominates at 97.9% of 105,484 rows, while '+' (present) appears in only 31 records (~0.03%), making positive-class detection extremely challenging. The near-zero entropy (0.147) confirms almost no informational variance in this column as-is.

Treatment: Flag severe class imbalance ('+' = 31 of 105,484); apply oversampling or class-weighted modelling, and consider binary encoding after collapsing '-' vs. non-'-'.

spreadGlottis categorical

constrictedGlottis categorical feature

This column captures a clinical or phonetic annotation for constricted glottis, encoded as presence/absence symbols across 105,484 records with no nulls and only 7 unique values. The dominant value is '-' (absent/negative), appearing in 94.5% of rows (99,727), while '+' (present) accounts for just 3,383 rows — indicating a rare positive condition. The compound values ('+,-', '-,+', '+,-,-', '-,-,+') suggest multi-observation or multi-segment sequences for a tiny fraction of records, but with counts of 141, 93, 1, and 1 respectively, these are near-negligible. The extremely low entropy (0.372, entropy ratio 0.132) confirms the column is heavily imbalanced toward the negative class.

Treatment: Binarize into presence/absence flag; treat compound multi-segment values as positive; account for severe class imbalance (94.5% negative) during modelling.

fortis categorical feature

This column is a three-valued categorical flag, likely representing a directional or strength indicator with values '-', '0', and '+'. The dominant value is '-' at 68.1% (71,867 rows), '0' accounts for 31.5% (33,202 rows), and '+' is strikingly rare at only 415 occurrences (~0.4%), creating a heavily imbalanced distribution. The near-absence of '+' values is a notable surprise and may indicate a rare positive condition, signal, or classification outcome worth investigating for class imbalance before modelling.

Treatment: Ordinal-encode as -1/0/1 or one-hot encode, and apply class-imbalance handling (e.g. oversampling '+') before modelling.

lenis categorical feature

This column represents a ternary linguistic or phonological feature, almost certainly a 'lenis' (weak/voiced) marker with values '-' (absent/irrelevant), '0' (neutral), and '+' (present/lenis). The dominant value is '-' at 68.1% (71,866 rows), while the positive lenis marker '+' is strikingly rare at only 416 occurrences (~0.4%), creating a heavily imbalanced distribution. No nulls exist across all 105,484 rows, and entropy is moderate at 0.93 bits, well below the theoretical maximum, confirming the skew toward the negative class.

Treatment: Encode as ordinal or one-hot; be aware that the '+' class (416 samples) is severely underrepresented and will require oversampling or class-weight adjustment before modelling.

raisedLarynxEjective categorical feature

This column encodes the presence or absence of a raised-larynx ejective phonological feature, using a diacritic notation system common in linguistic databases (e.g., IPA or feature matrices). The dominant value is '-' (absent/not applicable), appearing in 96.4% of 105,484 rows, creating severe class imbalance with an entropy ratio of only 0.10. Minority values include '0' (2,150 rows), '+' (1,573 rows), and compound sequences like '-,+' and '+,-', suggesting some segments receive multi-valued or composite feature annotations. The near-absence of nulls (0.0%) indicates complete annotation coverage, but the extreme skew means this feature will contribute negligible signal in most modelling contexts without deliberate resampling or grouping.

Treatment: Collapse compound values ('-,+', '+,-', '-,-,+') into a unified category and apply oversampling or class-weight adjustment before modelling given 96.4% imbalance.

loweredLarynxImplosive categorical feature

This column encodes a phonological feature — specifically whether a sound has a lowered larynx implosive articulation — using a symbolic notation system ('+', '-', and combinations). The distribution is severely imbalanced: 97.3% of the 105,484 rows carry the default/absent value '-', with only 2,150 instances of '0', 716 of '+', and fewer than 10 combined entries. The near-zero entropy ratio (0.088) confirms this column carries almost no information for most records, which would make it a very weak standalone predictor.

Treatment: Flag the severe imbalance; consider binarising ('+' vs. all others) or dropping if the rare positive class is insufficient for modelling purposes.

click categorical label

This column captures click direction or interaction type, encoded as symbolic tokens ('-', '0', '+') with combination values ('+,-', '-,+'). The dominant value is '-' at 68.2% of 105,484 rows, while positive-click signals ('+', '+,-', '-,+') account for only ~311 rows combined (~0.3%), suggesting a heavily imbalanced outcome or event flag. The presence of multi-value strings like '+,-' and '-,+' implies ordered sequences of click events were sometimes collapsed into a single cell rather than normalized — this warrants investigation before modelling.

Treatment: Ordinal-encode or one-hot after splitting multi-value entries ('+,-', '-,+'); treat class imbalance (~0.3% positive) before using as a classification target.