УРОВЕНЬ АКТИВАЦИИ И СПЕЦИАЛЬНЫЕ ОГРАНИЧЕНИЯ ПРИ МАТЕМАТИЧЕСКОМ МОДЕЛИРОВАНИИ РЕФЕРЕНЦИАЛЬНОГО ВЫБОРА

Abstract

The choice of a referential form for mentioning a referent in discourse is primarily guided by the current level of referent activation in the speaker's working memory (activation level, AL). AL, in its turn, is influenced, first, by a wide range of factors, and second, by referential filters, which are special contexts that pre-determine the use of a lexically full referring expression even in cases when the AL is sufficiently high for pronominalization. Examples of referential filters are the referential conflict filter and the world boundary filter. In the current article, we introduce yet another component in the model of referential choice (RC): RC constraints. Eg., in English, if the referring expression is syntactically a noun premodifier (the company shares), it assumes the form of a full noun phrase irrespective of the activation factors and the filters in action. Based on the WSJ MoRA 2020 corpus of English newspaper articles, we introduce three types of RC constraints: (a) noun premodifier position, (b) full noun phrases with specific functions, and (c) direct speech, and show these constraints to be highly frequent in the corpus, which suggests that they should be taken into account as part of RC models. The assumption that including such constraints in RC models may contribute to overall prediction accuracy is further supported by applying machine learning algorithms to RC modelling on corpus data.