Investigation of Cognitive Effort Distribution to Translation Subtasks by Eye Tracking and Computation Modeling

This project is proposed to conduct in-depth empirical research by means of eye tracking and computational modeling to address the following academic questions that have recently become hot spots in cognitive studies of translation process and human machine interactive translation (HMIT): How translation is carried out from the cognitive perspective? What cognitive activities are involved in the translation process? How translators' cognitive effort is distributed to cognitive activities and subtasks of translation, and which ones are more effort-demanding than the others? Exploring sound answers to these questions can not only deepen our understanding of the nature of translation but more importantly indicate (1) what computer aids are more demanded in alleviating translators’ cognitive workload and, accordingly, (2) how to design and implement HMIT for maximal utility in enhancing users' productivity and creativity. To explore these questions, we will first acquire translators’ motor event data (such as eye movements, keystrokes and mouse clicks) by eye tracking in three typical realistic scenarios: translation without vs. with a bilingual dictionary, and translation as postediting of state-of-the-art MT output with a range of computer aids (including translation memory and examples at various granularity levels). The data will first be used for computational modeling of time intervals between keystrokes during continuous typing (vs. not typing) of target texts, in order to infer sequential patterns of motor events corresponding to Jakobsen’s (2011) notion of micro-cycle, each of which is assumed to translate a text chunk (or termed translation unit) of a source sentence, and then for unsupervised machine learning to cluster subsequences of micro-cycles into clusters corresponding to cognitive activities for specific information-processing functions. In this way, we will arrive at computational models for essential cognitive activities in pure translation and in use of a dictionary and more sophisticated computer aids to facilitate the translation. Cognitive effort distribution will be computed in a comparative way according to machine learning results in terms of gaze duration and pupil size (a known indicative index of cognitive workload intensity). Based on the distribution, we can have reliable estimation of the effectiveness of various computer aids in reducing translators' workload. The research outcomes from this novel methodology are expected to disclose many unseen cognitive details of the translation process and also point HMIT to the most promising future direction, i.e., bettering computer aids for those translation subtasks that demand more cognitive effort.