Behavioural Translation Style Space (BTSS)

• BTSS is a framework that unifies latent style spaces for image translation with hierarchical behavioural spaces for human translation production.
• In image translation, BTSS employs losses like triplet-margin and style regularization to achieve compact, disentangled, and smooth latent manifolds, enhancing metrics such as PS and FID.
• For translation production, BTSS organizes keystroke and gaze data into multilayered behavioural templates, enabling quantitative simulation of affect, cognition, and translator style.

Searching arXiv for the provided BTSS papers to ground the article in current records. arxiv_search(query="Behavioural Translation Style Space", max_results=10) arxiv_search(query="Smoothing the Disentangled Latent Style Space for Unsupervised Image-to-Image Translation", max_results=5) arxiv_search(query="Toward a Behavioural Translation Style Space: Simulating the Temporal Dynamics of Affect, Behaviour, and Cognition in Human Translation Production", max_results=5) Behavioural Translation Style Space (BTSS) denotes, in the cited arXiv literature, two technically distinct constructions that share an emphasis on structured, gradual transitions in a latent or behavioural space. In unsupervised image-to-image translation, BTSS names the latent multi-domain style space in which style codes live and through which smooth, content-preserving semantic interpolation is sought (Liu et al., 2021). In translation-process research, BTSS denotes a hierarchical, multidimensional space of behavioural templates or patterns instantiated over time during human translation production, inferred from keystroke and eye-tracking data and used as the substrate for computational simulation (Carl et al., 16 Jul 2025). The term therefore refers not to a single standardized formalism, but to a family of space-based representations for translation behaviour in different domains.

1. Terminological scope and formal definitions

In the image-to-image setting, the relevant objects are a set of training images partitioned into style domains,

$\mathcal{X}=\bigcup_{k=1}^m \mathcal{X}_k,$

and a style-code space partitioned into domain-specific subsets,

$$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$

Each style code $s\in \mathcal{S}_j\subset \mathbb{R}^d$ is a $d$-dimensional vector that drives the appearance of the output in domain $j$. A translation is written

$$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$

In this formulation, the entire BTSS is the union of all $\mathcal{S}_j$, and the desired geometry is explicitly described as compact, disentangled, and smooth (Liu et al., 2021).

In the translation-production setting, BTSS is defined as a hierarchical, multidimensional space of behavioural templates or patterns. Let

$L=\{L_1,\ldots,L_6\}$

be the set of layers, with $L_1$ through $L_6$ corresponding to Activity Units, Keystroke Bursts, Production Units, Affective States, Translation Phases, and Translator Styles. If each layer $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$0 is associated with a measurable parameter space $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$1, then

$$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$2

Observed behavioural data $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$3 are mapped into this space by a feature-extraction function

$$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$4

with $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$5. The same construction is also described as a continuous manifold $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$6, where each coordinate axis corresponds to one numerical feature such as AU mean duration, log KBI, or RelDur_L (Carl et al., 16 Jul 2025).

Context	BTSS object	Primary data
Unsupervised image-to-image translation	Multi-domain latent style-code space $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$7	Images, domain labels, style codes
Human translation production	Hierarchical Cartesian product $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$8	Keystrokes and gaze

The two definitions are formally different. One is a learned style manifold for generative image translation; the other is a hierarchical behavioural state space for modelling human translation production.

2. Latent BTSS in unsupervised image-to-image translation

In the image-translation formulation, BTSS is the latent space in which style codes are generated, encoded, and traversed. Two mechanisms produce style codes. Under reference-style encoding,

$$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$9

and under random-style mapping,

$s\in \mathcal{S}_j\subset \mathbb{R}^d$0

The generator then translates a source image $s\in \mathcal{S}_j\subset \mathbb{R}^d$1 according to the target-domain style code $s\in \mathcal{S}_j\subset \mathbb{R}^d$2 through $s\in \mathcal{S}_j\subset \mathbb{R}^d$3 (Liu et al., 2021).

The intended geometry of this space is specified in three parts. It should be compact, meaning that there are no “holes” where the network never saw data; disentangled, meaning that style codes cluster by domain; and smooth, meaning that small moves in $s\in \mathcal{S}_j\subset \mathbb{R}^d$4 produce small, gradual changes in output. The paper summary states that vanilla GAN-based domain losses tend to push clusters far apart, which is beneficial for disentanglement but detrimental to smoothness, or allow codes to wander, which is detrimental to compactness. The BTSS objective is therefore to reshape a “clustered but gappy” encoding into a “compact, disentangled, and smooth” latent geography (Liu et al., 2021).

The architectural realization is generic rather than model-specific. A content encoder $s\in \mathcal{S}_j\subset \mathbb{R}^d$5 extracts content features from $s\in \mathcal{S}_j\subset \mathbb{R}^d$6; a style encoder $s\in \mathcal{S}_j\subset \mathbb{R}^d$7 maps a reference image to $s\in \mathcal{S}_j\subset \mathbb{R}^d$8; a mapping network $s\in \mathcal{S}_j\subset \mathbb{R}^d$9 maps random noise and a domain label to $d$0; a generator $d$1 injects $d$2 into the content features via AdaIN or style blocks; and a discriminator $d$3 with domain-specific heads judges real versus fake in each target domain. Exact block diagrams are inherited from base models such as StarGAN v2 or TUNIT, with the BTSS-shaping losses added on top (Liu et al., 2021).

3. Loss shaping, interpolation smoothness, and empirical behaviour

The latent BTSS is shaped by three auxiliary losses added to the base objective. The triplet-margin loss

$d$4

promotes domain-wise clustering with a controlled margin $d$5. It uses an anchor $d$6, a positive $d$7 from the same domain, and a negative $d$8 from a different domain. Its stated effect is to keep domains disentangled while preventing them from moving arbitrarily far apart, thereby shrinking inter-domain gaps (Liu et al., 2021).

The style-regularization loss

$d$9

encourages all style codes, including both $j$0 and $j$1 outputs, to remain near the origin. It is described as an approximation of a KL-to-$j$2 term from VAEs, simplified to a sample-based $j$3 penalty. Its effect is to pull style clusters inward and fill “voids” in BTSS. The perceptual-content loss

$j$4

penalizes LPIPS distance between the source image and its translation, so that movement in style space does not arbitrarily destroy input content. These terms combine as

$j$5

and the training objective becomes

$j$6

To evaluate whether interpolation in BTSS is gradual and predictable, the paper introduces the Perceptual Smoothness (PS) metric in $j$7. Given an interpolation path $j$8 between two style codes and its corresponding generated-image sequence $j$9, PS combines two terms. Alignment compares end-to-end LPIPS distance with the accumulated small-step LPIPS distances, and reaches $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$0 when the path is perfectly colinear in perceptual space. Uniformity compares the perceptual sizes of individual interpolation steps, and reaches $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$1 when all steps are equal. PS is the harmonic-mean smoothness score over these two quantities and penalizes “jumps” or “detours” in BTSS (Liu et al., 2021).

Representative results for the “gender” task on CelebA-HQ $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$2 compare StarGAN v2 with StarGAN v2 plus $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$3:

Model	PS $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$4	FID $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$5
StarGAN v2	0.272	48.35
Ours (w/ $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$6)	0.504	23.37

The same comparison also reports FRD decreasing from $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$7 to $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$8 and LPIPS changing from $$\hat{x}=G(x,s), \qquad x\in\mathcal{X}_i,\; s\in\mathcal{S}_j.$$9 to $\mathcal{S}_j$0. The summary states that PS nearly doubles, FID is halved, and FRD drops, indicating smoother transitions, higher overall image quality, and better content or identity preservation. Similar gains are reported across “smile” and “age” tasks, on the AFHQ animal dataset, and in the “truly unsupervised” setting with TUNIT. Qualitative interpolations are described as turning between-domain walks, such as male $\mathcal{S}_j$1 female, from artifact-prone trajectories into continuous morphological change (Liu et al., 2021).

4. Hierarchical BTSS for translation production

In human translation production, BTSS is organized into six embedded processing layers with distinct time scales and linked top-down and bottom-up interactions. The six layers are defined as follows (Carl et al., 16 Jul 2025).

1. Activity Units (AUs): uninterrupted stretches of gaze and/or keystrokes, represented as a sequence of observations

$\mathcal{S}_j$2

No additional state-transition model is imposed at this lowest level.

2. Keystroke Bursts (KBs): clusters of successive keystrokes separated by short pauses. Let $\mathcal{S}_j$3 be inter-keystroke intervals within words and $\mathcal{S}_j$4 intervals between words. Translator-specific thresholds are defined by

$\mathcal{S}_j$5

A KBI occurs when the pause is greater than $\mathcal{S}_j$6 but smaller than $\mathcal{S}_j$7; a Production Unit Break occurs when the pause is at least $\mathcal{S}_j$8. KBs are modelled as a discrete-time Markov chain over the states “typing” and “pause”.

1. Production Units (PUs): groups of KBs delimited by PUBs. They are treated as higher-level cognitive chunks and represented by a semi-Markov chain whose sojourn in state “$\mathcal{S}_j$9” lasts $L=\{L_1,\ldots,L_6\}$0.
2. Affective States (HOF): a discrete hidden-state process on Hesitation, Orientation, and Flow, represented as a POMDP factor

$L=\{L_1,\ldots,L_6\}$1

with $L=\{L_1,\ldots,L_6\}$2 and action, transition, observation, and reward components specified for the affective layer.

1. Translation Phases: orientation, drafting, and revision. Phase segmentation is deterministic:

$L=\{L_1,\ldots,L_6\}$3

2. Translator Styles: one of five clusters, “rapid,” “deliberate,” “confident,” “balanced,” and “cautious,” assigned by K-means on normalized features $L=\{L_1,\ldots,L_6\}$4.

The framework further posits continuous-time differential equations for three core layers, affect ($L=\{L_1,\ldots,L_6\}$5), behaviour ($L=\{L_1,\ldots,L_6\}$6), and cognition ($L=\{L_1,\ldots,L_6\}$7):

$L=\{L_1,\ldots,L_6\}$8

An explicit instance is also given:

$L=\{L_1,\ldots,L_6\}$9

$L_1$0

$L_1$1

Top-down and bottom-up interactions are mediated by precision weights $L_1$2 controlling the influence of prediction errors versus priors.

5. Behavioural signal encoding and computational simulation

The behavioural BTSS is grounded in time-stamped keystroke and gaze streams. The key and gaze channels are encoded as

$L_1$3

$L_1$4

At each millisecond within the current AU, these accumulate insertion keystrokes, deletion keystrokes, target-text words produced, total linear-reading time on ST/TT, total re-fixation time, and total scattered-fixation time. Preprocessing consists of four stated steps: log-transforming pause durations and burst lengths by $L_1$5; session-wise $L_1$6-normalization; outlier removal by clipping $L_1$7; and KBI/PUB threshold estimation through the median-based formulas above (Carl et al., 16 Jul 2025).

The computational translation agent is implemented as a hierarchical POMDP in the PyMDP library. Its A-layer module has hidden states $L_1$8 and a Gaussian observation model over $L_1$9. Its B-layer module includes the micro-actions “type,” “pause,” “look-ST,” and “look-TT,” with transitions conditioned on A-layer suggestions. Its C-layer module contains high-level planning states and produces PUB decisions and predictions of next-PU content. At each time step, the agent performs a Bayesian posterior update,

$L_6$0

computes expected free energy

$L_6$1

and selects the action minimizing expected free energy. Learning includes Dirichlet concentration updates for transition matrices, Variational Bayes for module parameters, and online style adaptation through prior updates over translator styles.

The experimental basis is CRITT TPR-DB v2023, comprising 521 English-to-{DE, NL, ES, DA, ET, PT, HI, AR, NO} sessions and 357,639 AUs automatically annotated with HOF labels by Random Forest. Translator styles are assigned by constrained K-means. Reported evaluation metrics and findings include the following (Carl et al., 16 Jul 2025).

Quantity	Reported value
HOF classification $L_6$2 on held-out sessions	$L_6$3
Translator-style clustering stability (ARI)	$L_6$4
PU alignment accuracy, agent vs. human	78%

The same results summary reports $L_6$5, an average DTW-distance of $L_6$6 between simulated and human time series, and phase-dependent HOF distributions in which Orientation corresponds to $L_6$7 O, Drafting to $L_6$8 F, and Revision to $L_6$9 H. AU occurrence rates and durations are also given, as are language-dependent threshold differences: Spanish has median $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$00 ms and $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$01 ms, whereas Hindi has $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$02 ms and $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$03 ms. Translator-style clusters differ in pause structure, with Style 0 (“rapid”) showing the lowest log PUB median at approximately $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$04 and Style 4 (“cautious”) the highest at approximately $$\mathcal{S}=\{\mathcal{S}_1,\ldots,\mathcal{S}_m\}.$$05.

6. Conceptual relations, misconceptions, and extensions

A common source of confusion is to treat BTSS as a single domain-independent construct. The cited literature does not support that interpretation. In one usage, BTSS is the learned multi-domain style latent space of a generator; in the other, it is a six-layer behavioural space that embeds affective, behavioural, cognitive, and stylistic regularities in human translation production (Liu et al., 2021, Carl et al., 16 Jul 2025).

The shared conceptual core is structural rather than formal. Both uses are concerned with trajectories through a space and with whether those trajectories are orderly, interpretable, and behaviourally meaningful. In the image case, the relevant regularities are gradual changes in generated images under interpolation, assessed by alignment and uniformity in perceptual space. In the translation-production case, the relevant regularities are temporal dynamics in gaze, keystrokes, affect, and planning, assessed by classification, clustering, alignment, correlation, and DTW-distance. This suggests a family resemblance between the two BTSS notions: each seeks a representation in which local movement corresponds to coherent change in observable behaviour.

The extension mechanisms also differ. In the image case, the BTSS framework is explicitly designed to be plugged into existing I2I approaches such as StarGAN v2 or TUNIT by adding the triplet-margin, style-regularization, and perceptual-content losses. In the translation-production case, the formalism is designed to support simulation and manipulation of translation styles under controlled priors, with proposed extensions to additional modalities such as mouse-clicks and revision keystrokes, social or contextual layers, continuous personalization, and generalization to simultaneous interpreting or other bilingual tasks through re-tuning POMDP observation and transition matrices (Liu et al., 2021, Carl et al., 16 Jul 2025).

Within these limits, BTSS is best understood as a representation-theoretic programme rather than a single model class. In one branch it denotes a behavioural latent geography for image morphing translations; in the other it denotes a behavioural state space for modelling the temporal dynamics of human translation production.