MaPPLexiC aims to fill the gap in modern language theories concerning the interpretation and description of one type of idiosyncratic word constructions that stands out in terms of its frequency of use and as a measure of language proficiency, namely lexical collocations, and refute the wide-spread assumption that their production is ad hoc and therefore cannot be described in systematic terms. Inspired by the dramatic advances in Deep Neural Networks (DNNs) MaPPLexiC considers a lexical collocation identification and classification DNN as a cognitive model, whose internal neuron activation vectors during the assessment of the collocation status of a given word combination can be translated into interpretable semantic, contextual, and socio-cultural features of the collocation elements and generalized into “collocation production principles” that dictate which features lexical items must possess to form a lexical collocation. In other words, what MaPPLexiC targets is a lexical collocation grammar. For this purpose, the Project will adapt and advance state-of-the-art techniques for the derivation of interpretable features from DNN’s activation vectors of individual lexical collocation samples and neural clustering techniques that will facilitate the generalization of the obtained “profiles” for the individual samples to more generic collocation production principles. Considering that collocation construction differs from language to language, the Project has a strong multilingual orientation. The investigation will be carried out on pairs of Germanic (English, German), Romance (French, Spanish), Finno-Ugric (Finnish, Hungarian), and Slavic (Czech, Russian) languages. The collocation profiles of translation-equivalent collocation samples and the language-specific collocation production principles will be contrasted with the goal to develop cross-language collocation production transfer techniques, which will highly benefit collocationally under-resourced languages.