Translation choices and their variability: The role of network structure
(2026) Translation in Transition conference — Location: RWTH Aachen University, Germany (9.September.2026)
Files
No attached file found for this publication.
Details
- Authors
- Abstract
- Halverson’s (2003, 2017, 2024) Gravitational Pull hypothesis proposes a cognitive explanation for source and target language influences in translation, in which translation choices are influenced by the organization of linguistic knowledge in networks (Langacker 2008, Diessel 2019), such that particularly salient connections exert gravitational pull toward the SL, magnetism toward the TL, or connectivity between the two. After an initial consideration of entrenchment, measured with corpus frequencies (Halverson 2017), subsequent work has considered additional aspects of salience, including recency and cross-linguistic similarity (Lefer & De Sutter 2022, Heilmann et al. 2022, Marco & Tello 2024). My research applies this model to understand the variability of translation choices, i.e. why different translators produce varied translations for some items but converge on one solution for others. Since variation in traditional parallel corpora is confounded by heterogeneous source contexts, I examine multiple same-text translations from the Multilingual Student Translation corpus (MUST; Granger & Lefer 2020), thus controlling for all factors apart from the translators themselves. The MUST texts selected here are specialized in sustainable finance and translated by UCLouvain Master students in Translation; these students represent an under-researched intermediate position between naïve bilinguals and professional translators. Unlike research in bilingualism and translation process (Tokowicz 2014, Schaeffer et al. 2016), where the presence of alternative translations is invoked to explain greater cognitive effort, I approach this variability as a phenomenon that itself requires cognitive and linguistic explanations. Malmkjaer has suggested that the use of multiple translation data might prove “particularly fruitful” where translator’s “varied realisations” differ most (1998: 539). Here, I hypothesize that a cognitive network with high entropy (having a large number of connections with evenly distributed salience) will give rise to a larger number of more evenly distributed translation types. I focus on French translations of English noun sequences in the specialized domain of sustainable finance. [N+N] structures are common in English (e.g. fossil fuel, investment decision) but pose difficulties for French translation (Lefer & De Clerck 2021) due to formal differences, the availability of synonyms, and ambiguous semantic relationships. These challenges contribute to translation variability. From 11 English MUST texts and 176 French translations, I have identified 122 [N+N] sequences and 4,660 translations. I start by describing the nature and extent of variability, both overall and within each of two or three construction slots corresponding to the dominant [N+Prep+N] or [N+A] forms in French. The latter can reveal how different choices combine to produce the observed variability. For instance, the translations of research firms (Figure 1) display onomasiological (different near-synonyms for firm) and semasiological patterns (different interpretations for the meaning of research and for the N-N semantic link). To address the question to what extent this variability can be explained by the structure of mental networks, I quantify two components of variability: the (normalized) number of translation types and the evenness of their distribution. As network structure is not directly observable, I operationalize the number and entrenchment of connections with a combination of corpus and experimental evidence (Figure 2), reflecting Schmid’s (2020) insight that usage is shaped by both collective and individual processes. Frequencies for nouns and noun sequences are obtained from English and French reference corpora for general and specialized language. These are supplemented with measures of association strength and with type frequencies and distributions, to provide more cognitively realistic analyses (Gries 2022). Cross-linguistic co-occurrence metrics are calculated from parallel reference corpora (Halverson 2017, Marco 2021). Psycholinguistic evidence consists of responses and reaction times (RT) in monolingual and bilingual elicitation and judgment tasks (Figure 2b). These experiments were conducted with 40 students who also contributed translations to the MUST corpus, thus permitting inter-individual comparisons of results. Results to date indicate that noun sequences with high frequency in specialized English tend to have fewer translation types and converge on a dominant type (Figure 3). Common terms like climate change have established translations (changement climatique), while rare combinations like product landscape lead to more diverse solutions. Frequencies in general English appear to have little effect; in fact, general and specialized frequencies are only weakly correlated, highlighting the importance of considering both. The experimental evidence for entrenchment points in the same direction: stimuli with shorter RTs are associated with significantly lower translation variability. In addition, in the monolingual elicitation tasks, stimuli with short RTs tended to generate more responses and vice versa. This provides internal validation for both RT and response number as measures of ease of recognition and production. Furthermore, RTs in all tasks tended to be shorter for stimuli with high (specialized) English frequency. A second question concerns the role of individual translators’ linguistic profiles. For the students who have contributed MUST texts and experiment data, several measures of language proficiency and translation experience are documented as metadata. Their effect on RTs in the bilingual elicitation experiment is nuanced: processing is faster with increased translation experience, as expected, but tends to slow down with increased English and French proficiency. The latter suggests that when a richer network is activated, a wider range of subtle distinctions needs to be evaluated. However, caution is indicated by the fact that, after fitting a LMM including these metadata, a large proportion of between-participant variation in RT remains unaccounted for (Figure 4). The effect of these individual differences on translation variability is also analyzed. In conclusion, results to date show an overall alignment between corpus and experimental evidence for entrenchment, and both support a significant effect of network structure on translation variability. Similar analyses of French reference corpus frequencies are expected to confirm that the number and proportional distribution of translation types is broadly congruent with attested numbers and relative frequencies. (References)
- Affiliations
Citations
Prinzie, T. (2026, September 10). Translation choices and their variability: The role of network structure. Translation in Transition conference, RWTH Aachen University, Germany. https://hdl.handle.net/2078.5/280368