Juan Pablo Vigneaux

Text-to-speech systems built on residual vector quantization often generate audio in two stages: a transformer backbone emits the first level of each frame’s token stack, and a smaller depth model generates the rest autoregressively. We ask which components of such a model are responsible for prosodic intonation, more precisely for generating the English nuclear tune (for instance, the falling pitch melody that distinguishes an assertion from a question’s rising melody), and whether interventions on those components can be used to control intonation. Prompt-based control mostly fails; only a final question mark is a reliable lever. Interchange interventions in CSM and S2 locate the tune: the backbone’s hidden state carries it to the waveform through the depth model, and the first level of the stack carries little or none of it. Steering the backbone state with a difference-of-means vector computed on new material generated with a final period or question mark moves the tune, but the push toward an assertion degrades the voice. Steering vectors computed instead from the backbone states associated with human imitations of theoretically-specified tunes, whose text carries no punctuation, and applied inside the generation loop, move most recordings to the other condition in both systems (even when fitted without the speaker being steered), at a small cost in predicted naturalness. These imitation-derived vectors are aligned with the one obtained from the two punctuated conditions far beyond a shuffled-label null, so the backbone state of a punctuated prompt already represents intonation. The steered recordings broadly become falling or rising, according to the intent, but the edge tone moves more coherently than the pitch accent, and the target tune cannot be selected precisely beyond this broad tendency.

Structural probes were introduced by Hewitt and Manning to reconstruct syntactic trees from a neural language model’s latent representations. They are evaluated by calculating the proportion of syntactic tree edges correctly reconstructed over an annotated corpus (as measured by undirected unlabeled attachment score). Here, we disaggregate this measure, considering undirected attachment score by label (UASL), which assesses the reconstruction accuracy of each syntactic relation separately, establishing important differences among relations that overlap linguistic distinctions. Moreover, we identify two factors that predict most of UASL’s variability across relations: (i) the mean and dispersion of the linear distance (on a log scale) between the related words, and (ii) the diversity (similarity-aware entropy) of the syntactic relation’s head. These results, which hold across a range of model sizes and architectures, shed light on the degree of abstraction of the representation of syntax in language models and the dependence of such representation on geometric properties of the embedding space.

We define \emphstratified measures on Euclidean space as convex combinations of rectifiable measures. They are possibly singular with respect to the Lebesgue measure and generalize discrete-continuous mixtures. A stratified measure ρcan thus be represented as \sum_i=1^k q_i \rho_i, where (q_1,..,q_k) is a probability vector and each \rho_i is absolutely continuous with respect to the m_i-Hausdorff measure \mu_i on a m_i-rectifiable set E_i (e.g. a smooth m_i-manifold if m_i>0 or a countable set if m_i=0). We introduce a set of strongly typical realizations of ρ^⊗n that occur with high probability; they are supported on a finite union of strata E_i_1\times ⋯\times E_i_n whose dimensions “per factor” concentrate around the \emphmean dimension \sum_i=1^k q_i m_i. For each n, an appropriate sum of Hausdorff measures on the different strata gives a notion of reference “volume”; the exponential growth rate of the typical set’s volume is quantified by Csiszar’s \emphgeneralized entropy of ρwith respect to μ=\sum_i=1^k \mu_i. This entropy satisfies a chain rule; the conditional term is related to the volume growth of the typical realizations in each stratum. Under suitable hypotheses, our notion of mean dimension coincides with Rényi’s information dimension when applied to stratified measures, but the generalized entropy used here differs from Rényi’s dimensional entropy.

We introduce the category of information structures, whose objects are suitable diagrams of measurable sets that encode the possible outputs of a given family of observables and their mutual relationships of refinement; they serve as mathematical models of contextuality in classical and quantum settings. Each information structure can be regarded as a ringed site with trivial topology; the structure ring is generated by the observables themselves and its multiplication corresponds to joint measurement. We extend Baudot and Bennequin’s definition of information cohomology to this setting, as a derived functor in the category of modules over the structure ring, and show explicitly that the bar construction gives a projective resolution in that category, recovering in this way the cochain complexes previously considered in the literature. Finally, we study the particular case of a one-parameter family of coefficients made of functions of probability distributions. The only 1-cocycles are Shannon entropy or Tsallis α-entropy, depending on the value of the parameter.