A psycholinguistic model for simultaneous translation, and proficiency assessment by automated acoustic analysis of discourse

Abstract

Two separate but related issues are addressed: how simultaneous translation (ST) works on a cognitive level and how such translation can be objectively assessed. Both of these issues are discussed in the light of a qualitative and quantitative analysis of a large corpus of recordings of ST and shadowing. Because the analysis of the act and the actuality of speaking simultaneously can reveal the workings of the processes which operate when speech is being generated, the knowledge gained from the analysis is used to characterise the cognitive processes concomitant with ST. The model formulated utilises this discourse-derived knowledge, many accepted facts in the psychology tradition, and evidence from controlled experiments that are carried our here. This model has three advantages: (i) it is based on analyses of extended spontaneous speech rather than word-, syllable-, or clause-bound stimuli; (ii) it draws equally on linguistic and psychological knowledge; and (iii) it adopts a non-traditional view of language called 'the linguistic construction of reality'. The model offers a realistic explanation of how ST works. The discourse-based knowledge is also used to develop three automated systems for the assessment of simultaneous translation: (i) content-based semi-automated; (ii) time structure-based; and (iii) coherence-based. For each system, several parameters of performance are identified, and they are correlated with assessments rendered by the traditional, subjective, qualitative method. The acoustic analysis of discourse, which uses standard signal processing techniques, leads to the conclusion that quality in simultaneous translation can be assessed quantitatively with varying degrees of automation. It identifies as measures of performance (i) three content-based standards; (ii) four time management parameters that reflect the influence of the source on the target language time structure; and (iii) two types of acoustical signal coherence. Proficiency in ST is shown to be directly related to coherence and speech rate but inversely related to omission and delay. High proficiency is also found to be associated with a high degree of simultaneity and prudence, but a low degree of time dissipation and inactivity. All the quantitative systems used here are shown to be independently capable of identifying the quality in translation, and their performance parameters capable of yielding congruent results regardless of whether the method used is fully or partly automated, and regardless of whether it is based on content, or on acoustical signals. This means that translation assessment does not need to be solely qualitative any more, and that these quantitative systems can be used to complement the traditional subjective qualitative method.