Efficient coding of different number notations from natural language: evidence from computational linguistics and behavioural data

The question whether numbers are represented independently from the notation or processed differently still sparks the debate. In two experiments, we employed the vector-space models derived data, informative about distributional patterns of both number words and Arabic numerals in the natural language to see whether the similarities and differences between notations can be connected with efficient encoding principles, strengthening the body of evidence from recent work. That is, we tested whether the specific distributional pattern of numerical information (i.e., number words vs. Arabic digits) in language can differently affect participants’ performance in comparison tasks. First, we compared how well the use of numbers in these two notations reflects the implied real numerical relationships between them. Next, we used a novel combinational approach to predict the performance of participants in two number comparison tasks, one with number words and one with Arabic digits, using the purely language-based data from each notation. We found that the linguistic proxy for distance is not only constituting better models for behavioural performance than the real numerical distance, but we also found evidence for a notation dependency. These observations suggest that the similar yet specific language regularities determine the way humans represent numbers, supporting in turn the notation reliant hypothesis for numbers representations.