Hesitation, orientation, and flow: A taxonomy for deep temporal translation architectures

Abstract

Numerous models have been proposed to describe and predict how human translations evolve in time. Some of these models suggest hierarchically embedded processes of fast and slow processing that unfold on different timelines. However, the assumed mental processes have often been conceptualized without a clear description of how they can be assessed, measured or retrieved in behavioral data. Other approaches suggest fragmenting behavioral data into various kinds of units, but the status of these units with respect to their cognitive reality is not always very clear.

In this paper, we propose a novel annotation taxonomy for behavioral data, assuming three broad states that translators experience during translation production: A state of orientation (O) reflects the experience of epistemic foraging in which a translator reads or scans a piece of the source text (ST) or searches for information. In a flow state (F), a translator engages in fluent translation production that is characterized by focus and involvement in the production process. A state of hesitation (H) can be described in terms of uncertainty or doubt which results in patterns of re-reading, text modification and disfluent translation production. This novel HOF taxonomy aims at describing mental states that elicit different experiential qualities during the translation process, associated with typical behavioral patterns that can be retrieved in recorded translation process data (TPD, i.e., logged keystrokes and gaze data).

We describe a manual annotation of a small set of TPD with the HOF taxonomy and we develop a method that ensures high inter-rater agreement (kappa 0.88). We show how our HOF translation states cluster into higher-level translation strategies (so-called policies). We discuss how these policies are optimized during the translation process and how translation states trigger off lower-level translation processes. We compare our taxonomy with other annotation approaches and propose a deep-temporal architecture that assumes a hierarchy of embedded translation processes that interact in various ways.