LINGUISTIC COMPETENCE THROUGH THE LENS OF GRAMMATICAL VARIATION. PART 1. CONCEPTUAL AND METHODOLOGICAL CONSIDERATIONS

Abstract

The linguistic competence is revealed empirically in the ability of native speakers to distinguish grammatical sentences from ungrammatical ones and to make judgments about acceptability of a certain language expression. This paper discusses the problem of language data consistency, which is examined with respect to comparison between the language competencies of a native speaker and a neural network.The high quality of modern neural network solutions for text and speech processing suggests that the corresponding neural systems have a form of a language competence. Since the individual characteristics of native speakers who are authors of texts collected in corpora are smoothed out when neural networks are trained, it is necessary to understand whether such a generalization can distort the linguistic reality. The paper aims to establish how the standard of language proficiency by native speakers can be revealed in acceptability judgments.In this study, we investigate the intralingual grammatical variation, which we consider to be a specific property of natural language. When speakers rate phenomena that display variation the expected consistency in their ratings decreases. In order to assess the consistency we propose two parameters. First, it is necessary to analyze the agreement of respondents within the linguistic community. We assume that within the linguistic community there are groups of speakers who have similar impressions on the acceptability of specific variants in the context of variation. Consequently, the ratings by the speakers from the same group are expected to vary in a rather small degree. Second, we consider the individual consistency within an experimental trial. As far as within one experiment the acceptability perception of a speaker does not change, we expect that variation within the scores for identical stimuli is restricted.In order to identify the possible range of judgments in case of variation we propose a taxonomy of phenomena that display variation considering two dimensions: the distribution of variants with respect to predictors and the relative acceptability of variants. We suggest taking two types of phenomena as the material for the study: (i) phenomena with little variation and a univocal choice of the variant determined by a predictor; (ii) phenomena with high variation when predictors are not defined or in the presence of several differently directed predictors. The two groups are represented by four phenomena that vary with respect to the expected distribution of variants and the type of grammatical interaction. The further quantitative analysis of experimental data on the suggested phenomena will provide a baseline for the analysis of variation within neural network architectures