Databases of grammatical properties of the world's languages are mostly empty. WALS is about 17% filled, and roughly 31% of languages in the widely used URIEL repository were silently given default values (URIEL+). Systems that fill these gaps rely on two kinds of evidence. One is inheritance: relatives share features. The other is global co-occurrence: features predict each other across all languages. Each works for some features and fails for others. A feature like basic word order tends to change rarely and stick within a lineage. A feature like a borrowed phoneme changes everywhere. Current imputers apply one kind of evidence uniformly to every feature.
We propose to measure, for each feature, how its variation is distributed over the family tree, using the same statistic that signal processing uses to find sparse transients in noise. Concretely, we split each family tree into nested binary splits and record, at each split, how much the feature differs between the two sides. These differences are the coefficients of a wavelet-like decomposition of the feature over the tree. If a feature is inherited, most splits show no difference and a few show a large jump, so the coefficients are heavy-tailed and their kurtosis is high. If a feature varies from one language to the next, the coefficients are spread evenly and kurtosis is near or below the Gaussian value. The hypothesis is that this number, computed from observed data alone, predicts how much a tree-structured predictor will beat a global low-rank imputer on that feature. If it does, kurtosis can set a per-feature mixing weight between the two.
Core Claim
The kurtosis of a feature's tree-contrast coefficients predicts which features benefit from phylogenetic back-off. This holds even under the SIGTYP 2020 protocol, which removes same-genus and nearby languages (SIGTYP 2020 Shared Task). Two further claims follow. First, the prediction is not reducible to the feature's frequency or its total variance across the tree. Second, kurtosis is a more stable choice rule than cross-validated model selection for sparsely observed features, where held-out estimates are noisy.
The novelty sits in the per-feature decision. Kurtosis is not used as a filter on raw columns. That use would be trivial, because the kurtosis of a binary variable is a function of its frequency alone. It is used as a diagnostic on a hierarchical binary decomposition of the tree.
Formally, for feature $f$ and internal binary node $n$ with observed left and right descendant sets of sizes $m_L, m_R$ and observed feature means $\bar{y}_L, \bar{y}_R$, the contrast is
$$d_{f,n} = \sqrt{\frac{m_L m_R}{m_L + m_R}}\,(\bar{y}_L - \bar{y}_R).$$
This is a Haar coefficient on the tree, and the scaling makes coefficients comparable across clade sizes. The diagnostic is the excess kurtosis $K_f = E[d^4]/E[d^2]^2 - 3$ over all nodes with $m_L, m_R \ge 2$. It is used as a mixing weight $w_f = \sigma(a(K_f - b))$ between a tree back-off predictor and a global imputer. Here $\sigma$ is the logistic function, and $a, b$ are two scalars fitted once on training features.
Relation to Prior Work
The kurtosis mechanism comes from signal processing. There, kurtosis is the standard criterion for picking the components of a decomposition that carry sparse, impulsive structure rather than broadband noise (Kurtosis Based Signal Component Selection). The closest template is threshold selection for UWB TOA estimation, which computes kurtosis over a dyadic wavelet tree of a radio pulse to separate the direct path from multipath clutter. We transplant that move from a time axis to a phylogeny.
That literature also constrains the design. Because kurtosis uses fourth powers, one spurious spike inflates it, which is why mode-selection work adds masking or ensembling. A single miscoded language would do the same to our coefficients. We therefore require minimum clade sizes and report a leave-one-clade-out robust kurtosis. Two prior uses of kurtosis on discrete or neural data, label-set optimization via activation kurtosis and KurTail, compute it over continuous activations. Neither touches categorical matrices with missing cells.
The hierarchical decomposition comes from Phylo-HS. It binarizes a multifurcating taxonomy by balancing child frequencies and routes each prediction through binary node decisions. We reuse its binarization to turn Glottolog trees into the binary splits on which contrasts are defined. Our goal differs: Phylo-HS assumes complete labels and seeks speed, while we use the decomposition to diagnose features under missing data.
The closest existing work comes from SIGTYP 2020 entrants that already use the tree. CrossLingference modeled feature change along lineages with continuous-time Markov processes, NEMO predicted from macro-family down to genus, and Panlingua used hierarchical SVMs that back off from branch to clade (SIGTYP 2020). All three apply the tree to every feature in the same way. We ask, per feature and before seeing held-out data, whether the tree should be trusted at all.
On the global side, Bjerva et al.'s collaborative-filtering model and URIEL+'s finding that SoftImpute beats autoencoders treat all features alike under one low-rank model. data2lang2vec trains one independent random forest per feature. That paper is also the motivating evidence: family features drove 75% of its predictions. This is strong evidence that inheritance matters, but it does not say for which features.
Two further points shape the design. Grambank deliberately avoids logical dependencies between features. That suits our per-feature diagnostic, which needs no implicational structure. Grambank's binary coding and more than 70% density also make the contrasts well estimated. Finally, the SIGTYP finding that the share of a target language's features that are missing is the main predictor of failure motivates evaluating separately by coverage band.
Research Program
We first build the decomposition. Glottolog family trees, as distributed through lang2vec, are binarized with Phylo-HS-style balancing. We also run three alternative binarizations to check that $K_f$ is not an artifact of arbitrary split order.
We then compute $K_f$ for every feature in Grambank and in the SIGTYP WALS subset, pooling contrasts within and across families. As a descriptive check, we test whether $K_f$ ranks features in a linguistically sensible way. We expect word-order and morphological-alignment features to score high and phoneme-inventory features low. This is a sanity check, not evidence for the claim.
Next we define the two predictors to be mixed. The tree predictor descends the binary decomposition. It predicts a held-out language's value from the deepest ancestral clade that still contains enough observed, protocol-eligible languages, which above genus level means family-level back-off. The global predictor is SoftImpute on the full language-by-feature matrix. The mixture uses $w_f$, with $a, b$ fit on a training split of features and then frozen.
The central analysis regresses the per-feature accuracy gain of tree over global prediction on $K_f$. The regression controls for feature frequency, number of observations, total contrast energy $E[d^2]$, and family count. This tests whether the fourth moment carries information beyond the second moment and simple rarity.
Baselines
The naive reference points are the SIGTYP frequency baseline, which reports macro-averaged accuracy of 0.49, and k-nearest neighbours at 0.48, both under the constrained protocol (SIGTYP 2020). We reimplement both on our exact splits so that the numbers are directly comparable.
The strongest reported system is ÚFAL, the constrained-track winner, at 0.75 macro-averaged accuracy. It combines neural networks over language embeddings with feature-correlation estimates. We compare on the official SIGTYP test split, where that number applies. Matching 0.75 with a two-parameter gate over two simple predictors would already be notable. Beating it by a margin larger than across-seed variation on the bootstrap would count as a win.
The strongest global imputer is SoftImpute as recommended by URIEL+. For it we found no reported result on Lacuna in SIGTYP-style accuracy. We run it with the same rank and regularization search budget we give our own components.
The strongest per-feature learned competitor is a data2lang2vec-style random forest per feature, with family, geography and neighbour values as inputs. Its exact numbers are truncated in the record, so we retrain it on our splits with an equal hyperparameter budget.
Three ablations isolate the key ingredient. The first replaces $K_f$ with second-moment energy $E[d^2]$, which tests the fourth moment specifically. The second replaces it with alternative sparsity measures of the coefficients, namely the Gini index and the L1/L2 ratio, which tests whether kurtosis is special or any sparsity measure works. The third replaces the per-feature weight with a single global mixing weight, which tests whether per-feature selection matters at all.
Why not just…?
Why not just pick the better predictor per feature by cross-validation? This is the most serious threat. With ample data it should match any diagnostic. Our bet is about the sparse regime. For features observed in a few hundred languages, which is WALS's average, cross-validated gains estimated under genus exclusion rest on few held-out cells and are noisy. $K_f$ instead pools structure from every node of the tree. We will compare kurtosis gating with cross-validation selection, stratified by number of observations. If cross-validation is equal or better in every stratum, we will report kurtosis only as an interpretable, label-free diagnostic and drop the selection claim.
Why not just use feature frequency or a standard phylogenetic-signal score? For a binary column, raw kurtosis is a function of frequency, so a naive version of this idea collapses to rarity. That is why we compute kurtosis on tree contrasts and control for frequency and contrast energy in the regression. If $K_f$ adds no explained variance beyond those controls, the claim fails.
Why not just give a per-feature random forest the family identity and let it learn when to use it? data2lang2vec does essentially this, and family inputs dominated its predictions. That suggests the learner already discovers inheritance where it exists. Under genus exclusion, however, the learner must infer family-level reliability from held-out-poor data. The gate supplies that reliability from the tree structure. We test this directly by adding $K_f$ as an input to the forest and checking whether it changes accuracy. A null result here would mean the forest already captures what the gate offers.
Proposed Experiments
We run on three datasets. On the SIGTYP 2020 WALS subset we use the official constrained split. On Grambank we use repeated random masking of 20% of observed cells, applying the same exclusion of same-genus languages and languages within 1,000 km. On URIEL+ we additionally check whether gated imputation changes LANGRANK transfer-selection quality. URIEL+ reports gains of 7.42% for POS and 6.82% for MT from imputed vectors over the original repository, and we test whether gating moves those further.
Metrics are macro-averaged accuracy over features, reported overall and by target-language coverage band and per-feature observation count. We also report the rank correlation between $K_f$ and the per-feature gain of tree over global prediction.
The decisive checks are four. First, $K_f$ must predict that gain on held-out features with a partial correlation that survives the frequency and energy controls. Second, it must beat the second-moment ablation. Third, it must match or beat cross-validation selection on low-observation features. Fourth, it must be stable across the four binarizations. If $K_f$ fails the first check, the idea is wrong. If it passes the first but loses to cross-validation everywhere, it survives only as a descriptive tool for typologists asking which features behave as inherited traits.
References
- URIEL+: Enhancing Linguistic Inclusion and Usability in a Typological and Multilingual Knowledge Base
- SIGTYP 2020 Shared Task: Prediction of Typological Features
- An alternative switching criterion for independent component analysis (ICA)
- Threshold selection method for UWB TOA estimation based on wavelet decomposition and kurtosis analysis
- Label Set Optimization via Activation Distribution Kurtosis for Zero-shot Classification with Generative Models
- KURTAIL : KURTOSIS-BASED LLM QUANTIZATION
- Phylo-HS: A phylogenetic hierarchical softmax for taxonomic classification
- A Probabilistic Generative Model of Linguistic Typology
- data2lang2vec: Data Driven Typological Features Completion
- Grambank’s Typological Advances Support Computational Research on Diverse Languages