Temporal Dynamics of Emotion and Cognition in Human Translation: Integrating the Task Segment Framework and the HOF Taxonomy

1 Introduction

This article introduces a novel generative framework for modelling the temporal dynamics of translation behavior. It posits that translation emerges from the interaction of three embedded layers of mental processing: (A) affective/emotional states, (B) behavioral, automated translation routines, and (C) cognitive, reflective thought. This ABC model of the translating mind suggests that these embedded processes can be inferred from observable behavioral data – primarily keystrokes and gaze patterns – captured through keylogging and eye-tracking technologies. Such data have long been foundational in Translation Process Research (TPR) and Cognitive Translation and Interpretation Studies (CTIS) for analyzing cognitive effort, translator expertise, as well as the behavioral impacts of text difficulty, translation quality, and others.

Among the methods used in TPR, pause analysis – the measurement of inter-keystroke intervals (IKIs) – has served as a proxy for translation effort (e.g., Carl, Schaeffer, and Bangalore 2016; Lacruz and Shreve 2014; Vieira 2017). However, defining and interpreting pause lengths remains contested (Couto-Vale 2017; Kumpulainen 2015). In this article, I propose a reinterpretation of pause analysis through the lens of the ABC model by integrating two conceptual tools: the Task Segment Framework (Muñoz and Apfelthaler 2022) and the HOF taxonomy (Carl et al. 2024).

The TSF, introduced by Muñoz and Apfelthaler (2022), distinguishes between Tasks – automated/routinized actions such as brief keystroke bursts – and Task Segments (TSs), which are longer, intentional production units composed of a sequence of one or more Tasks. These Task Segments are separated by Task Segment Pauses (TSPs), whose duration varies by translator and is thought to reflect moments of increased cognitive demand.^[1]

Complementing this, Carl et al. 2024 proposed the HOF taxonomy, which categorizes translation processes into three experiential states: Hesitation (linked to uncertainty and elevated cognitive effort), Orientation (characterized by attentional shifts and evaluative engagement), and Flow (a state of focused, effortless production often accompanied by positive affect). These affective states modulate different phases of attention and readiness for action, and are seen as responses to dynamic cognitive demands.

Both the TSF and the HOF taxonomy share a generative perspective: that internal translation processes – whether automated, reflective, or affective – leave discernible traces in behavioral data, such as IKIs structures and gaze trajectories. On this view, hidden mental states generate observable actions, and the structure of those actions reflects the underlying processes that gave rise to them. This reciprocal relationship resonates with Bayesian theories of cognition, including Predictive Processing (Seth 2021; Clark 2023) and Active Inference (Friston et al. 2023; Parr et al. 2022; Pezzulo et al. 2024), which conceptualize the human mind as operating on generative models that iteratively approximate and adapt to environmental conditions.

In light of these theoretical developments, Carl (2024) and Carl et al. (in print) propose simulating translation behavior via an artificial Active Inference agent organized into three interconnected ABC processing layers. The A-layer captures affective states and emotional responses; the B-layer reflects automated, routinized behavior; and the C-layer governs reflective and deliberate cognitive processes. Within this model, the TSF can be used to differentiate automated behavioral routines (B) from cognitive reflective behavior (C) based on pause structure, while the HOF taxonomy provides insight into the emotional states accompanying translation activity (A). Together, they support a multi-layered, embedded architecture in which translation processes unfold simultaneously across multiple cognitive-affective dimensions and timelines.

This article elaborates on that architecture by examining the relationships between Tasks, Task Segments, and HOF states using a large corpus of translation process data. Section 2 situates the proposed framework in relation to prior translation theories, highlighting its novel contribution. Section 3 outlines the TSF and HOF taxonomy and illustrates their segmentation logic using annotated progression graphs. Section 4 presents an empirical analysis of IKI distributions across 100 from-scratch translation sessions and explores the pausing patterns that define TSF segments. Section 5 offers conclusions in the light of Robinson’s (2023) ideosomatic theory, underscoring the integrative value of this model for future research in translation cognition.

2 Models of the Translation Process

Over the past 60 years, numerous approaches have emerged to explain and model translation processes in both humans and machines (MT). Despite fundamental differences, there are notable overlaps between these two approaches. This section provides a very brief review and situates the current proposal within this historical context.

Rule-based MT (RBMT), dominant until the early 1990s, modeled translation as a sequence of linguistic analyses: analysis, transfer, and generation. In Cognitive Translation and Interpreting Studies (CTIS), this mirrors the comprehension-transfer-production model (Angelone 2010), involving problem recognition, solution proposal, and evaluation. RBMT emphasized source text (ST) disambiguation, assuming transfer and target language generation to be relatively straightforward (Hutchins and Somers 1992).

The emergence of statistical MT (SMT) in the 1990s marked a shift. SMT built probabilistic models capable of generating many potential translations and selecting the most likely one using advanced decoding techniques. A similar shift occurred in Translation Studies, moving from equivalence-based linguistic theories (e.g. Nida 1964; Catford 1965) to Functionalism and Skopos Theory, which emphasized target audience needs and communicative purpose (Nord 2006; Pym 2003).

From the mid-1980s, cognitive theories were increasingly used in translation studies, focusing on the process of translation. Gile’s Effort Model (Gile 1985, 2009) highlight cognitive load and resource allocation in interpreting. Think-Aloud Protocols (TAPs) and later keystroke logging (Jakobsen 1999) offered empirical insights into mental operations, revisions, and hesitation patterns. Eye-tracking, introduced around 2005 in TPR, further revealed attention distribution and cognitive effort (Alves and Albir 2025).

Cognitive translation research has drawn from bilingualism and cognitive psychology. Theories like the Revised Hierarchical Model (Kroll and Stewart 1994) and the Bilingual Interactive Activation Model (Dijkstra and Van Heuven 2002) argue for non-selective language activation – bilinguals access all known languages simultaneously, a trait shared by multilingual large language models (LLMs), which also handle code-switching with ease.

Translation Process Research (TPR) has adopted dual-process theories, distinguishing between fast, intuitive (Type 1) and slow, deliberate (Type 2) processes (Kahneman 2011). Königs (1987), Hönig (1991), Lörscher (1991), Schaeffer and Carl (2013), Carl and Schaeffer (2017, 2019) and others made distinctions between automatized and strategic translation behaviors, under a variety of different nominations see (Alves and Albir 2025) for a recent overview.

The default-interventionist model of reasoning and decision-making (Evans and Stanovich 2013) posits that intuitive processing is the default procedure, while intervention only take place when difficulty or novelty demands attention. This idea is echoed across translation scholarship (e.g. Carl and Dragsted 2012; Dimitrova 2005; Tirkkonen-Condit 2005), and the Monitor Model (Schaeffer and Carl 2013), which assumes that automated routines are overseen by higher-level monitoring processes.

Robinson (2023), building on Peirce (1992), introduces the concept “emotional interpretant,” arguing that emotional awareness precedes conscious reasoning. Emotions influence cognitive load, attention, and decision-making (Hubscher-Davidson 2017), and shape how translators interpret and resolve ambiguity. Because emotional and cognitive states surface in observable behaviors – such as keystrokes or gaze – they may offer clues into translators’ internal states. The next sections explore taxonomies to identify automated, reflective, and affective mental processes in translation process data.

Figure 1:

A progression graph of a small snippet of the translation session (BML12/P03 T5). The graph represents a segment of approximately 40 s (116,000–156,000 ms) of an English-to-Spanish translation. The vertical axis plots to the ST on the left side and the TT on the right side. ST and TT words are aligned on a word or phrase level. A single ST word that maps into TT phrase is marked as a multi-word unit on the right vertical axis, such as “Increasing → Una_mayor”. An ST multi-word phrase that is translated into a single word appear as TT repetitions, such as “different from their own → otras otras otras otras”. Blue dots and green diamonds indicate eye movements on the ST and TT respectively. The black and red characters are insertion and deletion respectively. Activity Units (AUs) fragment behavioral translation data into six categories, as marked in colored boxes at the bottom of the graph. The type (color) of the AU determines whether the translator is involved in reading the ST or TT, translation production, or simultaneously reading and writing (see Table 1). TSs are marked as gray boxes in the top of the graph. The rectangular box in the middle of the graph (around time 136,000–140,000) is reproduced in Figure 2 with a finer-grained segmentation of the TS into six (sub) Tasks.

Figure 2:

This Progression graph shows a segment of approximately 4 s (136,000–140,000) in which an English segment “in the increasing” is translated into Spanish “un augmento en la”. The duration of a TS is indicated as a grey bar at the top, which is preceded and followed by a TSP (violet boxes). The TS consists of two successive Tasks which are separated by an RSP. Each Task contains one or more insertion and/or deletion keystroke(s). In Task 1 the translator produces “un a yum”. The last tree letters “yum” are presumably a typo, because these letters are successively deleted again in Task 2 (deletion “muy” is inverse order of “yum”). Task 2 then shows the correct continuation “ugmento” (to produce “augmento”) and the production of “en la”. As indicated by the “Fixation Units”, i.e., the blue and green striped boxes on the top, while the eyes of the translator fixate in the beginning and the end of this segment at several ST words (blue boxes are ST fixations), they move to the target window during the correction or the typo (green boxes are TT fixations). The Sequence of colored AUs at the bottom of the graph indicates the coordination of reading and writing behaviour (see Table 1). Note that Task 2 consists of two AUs of Type 6 and 5 which involve concurrent writing and reading of the TT and ST, respectively.

3 Fragmenting Behavioral Translation Data

The temporal structure of translation has been central to Translation Process Research (TPR) since the 1980s (e.g., Königs 1987, see also Alves and Albir 2025). With the advent of keylogging tools like Translog (Jakobsen and Schou 1999) and InputLog (Leijten and Van Waes 2013), TPR has become increasingly technology-driven. With these technologies, it can be observed how translators alternate between rapid, automatic processing and more deliberate, reflective behavior. Interkey intervals (IKIs) can be measured to offer insights into the hidden cognitive states, and when combined with eye-tracking data (Carl 2012; Carl, Schaeffer, and Bangalore 2016; Hvelplund 2016), they reveal not only how translations are produced, but also what information translators take in.

This section discusses segmentation methods of behavioral translation data that aligne with the three ABC layers of the translating mind. The HOF taxonomy, mainly based on gaze data, identifies affective states (A), while the TSF defines IKI thresholds to distinguish routinized processes (B) from reflective, cognitive processes (C).

4 An Empirical Investigation

This section presents an empirical analysis of the TSF and the HOF taxonomy based on a fragment of the CRITT TPR-DB (Carl, Schaeffer, and Bangalore 2016). It is partially a summary of data discussed in (Carl 2024) and (Carl et al. in print), here presented within a different perspective.

4.5 Relating RSPs and TSPs

Figure 3 shows the distribution of RSPs on the left and TSPs on the right for 32 Spanish and 22 Arabic translators, respectively. As is the case for all IKIs, also RSPs and TSPs show larger variability for Arabic, i.e., a flatter distribution, than Spanish. However, all RSPs are – for every translator – shorter than their TSPs.

Figure 3:

Distribution of RSPs (left) and TSPs (right) for Spanish and Arabic translators.

Figure 4 shows that RSPs tend to correlate with TSPs. This correlation is significant for Spanish (Spearman τ:0.68, p < 0.0001), while it is not significant for Arabic (Spearman τ:0.40 p:0.065).

Figure 4:

Correlation of RSPs and TSPs.

Interestingly, as plotted in Figure 5, there is a strong correlation between the number of Tasks (as delimited by RSPs) within a TS and the number of keystrokes produced in that TS (τ:0.74 and τ:0.73, p:0.000 for Arabic and Spanish respectively). While, on average, Arabic and Spanish translators engage in the same number of 2.2 Tasks per TS, Arabic translators show a larger variation (between min:1.2 and max:3.9, median:2.1) than Spanish translators (between min:2.1 and max:3.4, median 1.94).

Figure 5:

Correlation of total number of keystrokes per task segment and number of subtasks.

Spanish translators also produce more keystrokes per TS than Arabic translators do. A Spanish TS contains between 8 and 18 keystrokes (mean 11.2) while an Arabic TS has between 5 and 16 keystrokes (mean 9.4). A Spanish Task has between 3.9 and 7.4 keystrokes (mean 5.3) while an Arabic Task has between 2.7 and 6.0 keystrokes (mean 4.3).

We also observe a slightly negative effect of TS length on the number of keystrokes produced per Task: As the number of Tasks per TS increases, the number of keystrokes per Task decreases. This effect is significant for Spanish (τ:−0.52, p:0.002) but not for Arabic (τ:−0.18, p:0.41), which may have to do with the larger variability in the data and the smaller number of observations for our Arabic data set.

4.6 Types of Tasks

Following Muñoz and Apfelthaler (2022), we distinguish between three types of Tasks that involve different types of keystrokes:^[5] an insertion Task, A, has only insertion keystrokes (corresponds to Muñoz and Apfelthaler’s ADD), a deletion Task, D (not considered in Muñoz and Apfelthaler) has only deletions, and a change Task, C, has insertions and deletions (corresponds to Muñoz and Apfelthaler’s CHANGE). We omit the SEARCH Task, since in our translation sessions we have no external research.^[6]

Figure 6 shows average duration and keystrokes for the three Tasks, for Spanish and Arabic. The figure shows that there are systematic differences between the Spanish and the Arabic tasks. It can be expected that differences exist also between individual translators, and presumably also for different text types and translation goals. We showed that IKI profiles seem to be typical for specific translators, so that translators can be recognized (to some extent) by their IKI distributions. Figure 6 shows that on average all types of Tasks A, D, and C have more keystrokes for Spanish as compared to Arabic and the average duration is longer for Arabic than for Spanish.

Figure 6:

Number of keystrokes (left) and duration (right) for the three types of Arabic and Spanish Tasks. There are more keystrokes and shorter timespans in the Spanish data.

5 Discussion and Conclusion

This article proposes a hierarchically embedded model for simulating the translating mind, composed of three interacting layers: (A) an affective layer addressing emotional states, (B) a behavioral layer for automated routines, and (C) a cognitive layer simulating reflective thought. The model integrates two taxonomies: the HOF taxonomy, which categorizes affective translation states, and the TSF, which identifies behavioral markers of routine and reflective processing. Using these frameworks, the article analyzes keystroke and eye-tracking data of 100 Arabic and Spanish translation sessions to scritinize these ABC-layered processes.

Human translation production is structured into sequences of motor programs and Tasks – that is, actions such as adding, modifying, or deleting text. Tasks are separated by short pauses (respites, RSPs), while Task Segments (TSs) are separated by longer Task Segment Pauses (TSPs). Following Muñoz and Apfelthaler (2022), RSPs indicate involuntary breaks (which we locate here on the B-layer), while TSPs reflect conscious, planned pauses (i.e., C-layer units). Tasks and TSs thus signal routine and reflective behavior respectively, which depend on the typing speed of the translator.

The study shows that translators exhibit individual, but consistent pausing patterns suggesting personal, recognizable translation styles. While typing/pausing styles differ widely across translators, the structure of HOF states (A-layer units) seems surprisingly consistent across translators.

Thus, although typing speed may vary significantly between Arabic and Spanish translators, the frequency and type of translation actions (e.g., insertions, deletions) are remarkably similar. This suggests that while the “how” of translation (its temporal structure) is highly individual, the “what” (the nature of changes made) remains relatively consistent across languages and translators.

The proposed ABC architecture of the translation mind resonates with Robinson’s (1991, 2023) ideosomatic theory of translation, which views translation as embodied, affective, and cognitive. This ideosomatic theory emphasizes the translator’s bodily and emotional involvement, in addition to cognition and reasoning. The term ideosomatic combines the cognitive (ideo) and the physical (somatic), positing that translators’ decisions are often driven by intuitive, affect-laden reactions to text – e.g., what “feels right” – rather than purely analytic reasoning.

Robinson (2023) draws from Peirce’s (1992) triadic model of interpretants – emotional, energetic, and logical – which closely align with the ABC layers:

1. Emotional interpretant (A): the immediate, intuitive feeling response that shapes attention and meaning.

2. Energetic interpretant (B): embodied or motor reactions, expressed in typing and gaze behavior.

3. Logical interpretant (C): reflective reasoning that regulates and reinterprets emotional input.

Robinson argues that translation involves a “feeling-becoming-thinking” process: smooth translation arises when emotional and motor states align, but reflective thought may be called upon when there’s a mismatch. Though feelings and thoughts are internal, Robinson insists they are also transcranial – partly observable through social and sensorimotor behavior, including gaze and typing patterns. The internal processes can thus be traced through behavioral data.

This study builds on Robinson’s ideosomatic framework analyzing how the sequencing of HOF states, Tasks, and Task Segments reflects these layered processes. While the temporal structure of translation seems to be highly individual, the underlying actions and affective states remains relatively consistent – suggesting that shared cognitive-emotional mechanisms are common across translators. These findings, in combination with the ideosomatic theory, offer a nuanced understanding of how individual styles emerge from common/shared mental architectures.

Future research may investigate how the mental layers underpinning emotional states like Flow, Hesitation, and Orientation relate to cognitive concepts, such as attention shifts, working memory load, or retrieval processes. A more granular analysis of how different task types (ADD, DELETE, CHANGE) correlate with cognitive demands and emotional responses may also yield valuable insights, especially for training and process optimization. Moreover, fuller integration of eye-tracking data – through analysis of gaze fixations and saccadic patterns – could reveal how visual attention mediates emotional states and task complexity. Finally, translating these findings into practical guidelines for translator training and the development of emotion-sensitive translation tools would increase the model’s real-world impact, enabling more adaptive and cognitively informed approaches to translation practice and pedagogy.