A Probabilistic Approach to Language Structure

The translation of international legal instruments requires a high degree of accuracy and consistency. With the increasing demand for multilingual texts, translation memory tools and research on parallel corpora have proved to be particularly useful for the translation of repetitive documents, as well as for those subject to an evolutive drafting process and production. Moving from this assumption, the present study aims at predicting translation equivalents with the help of a probabilistic approach. Data (1.404.723 words) consists of a multilingual parallel corpus in four languages: English, French, German and Italian. All the documents have been taken from the EU secondary legislation and include Regulations, Decisions, Directives and Recommendations, chosen between the years 2001-04. Texts are all strictly ‘normative’ and discourse is expected to be precise with minimum scope for ambiguity. The main focus is prescriptive statements, namely deontic norms (permission, obligation, prohibition) and constitutive performatives. Their formulation is highly standardized in English both within and outside the EU context. On the other hand, their expression in other languages is more vague and extensive, with potential consequences on the translation of norms. Bearing these remarks in mind, our objective is: 1) to evaluate the degree of prescriptive standardization with reference to English and the other three languages, and 2) to predict translation equivalents in the other languages under the condition that (i) English legal drafting is highly standardized, (ii) the EU and the main English drafting guidelines tend to use modal verbs in prescriptive statements (iii) text types under examination are repetitive and reusable (iv) the four EU instruments can be more or less binding. English modals are used as the main entry point and entropy analysis is exploited to measure the number of alternatives (degree of uncertainty) occurring in the other three languages. By adding knowledge to a system (e.g. a more standardized formulation), one reduces the number of alternatives (uncertainty), which leads to a decrease of entropy and to a gain of information in the expression of the norm. Although language phenomena cannot be fully described, the results of this analysis have empirically proved that given a set of conditions, certain linguistic structures are more easily predictable than other when comparing several languages. These types of analysis can foster research in language testing, evaluation, and in the development of automated translation’s tools.