Discourseology of Linguistic Consciousness: Neural Network Modeling of Some Structural and Semantic Relationships Дискурсологія мовної свідомості: нейромережеве моделювання деяких структурних і семантичних взаємозв’зків

Objective. Study of the validity and reliability of the discourse approach for the psycholinguistic understanding of the nature, structure, and features of the linguistic consciousness functioning. Materials & Methods. This paper analyzes artificial neural network models built on the corpus of texts, which were obtained in the process of experimental research of the coronavirus quarantine concept as a new category of linguistic consciousness. The methodology of feedforward artificial neural networks (multilayer perceptron) Дискурсологія мовної свідомості: нейромережеве моделювання. . . 194 ©  Шимко  Віталій was used in order to assess the possibility of predicting the leading texts semantics based on the discourses ranks and their place in the respective linear sequence. Same baseline parameters were used to predict respondents’ self-assessments of changes in their psychological well-being and in daily life routine during the quarantine, as well as to predict their preferences of the quarantine strategies. The study relied on basic ideas about discourse as a meaning constituted by the dispersion of other meanings (Foucault). The same dispersion mechanism realizes itself in interdiscourse interaction, forming a discursive formation at a higher level. The method of T-units (Hunt) was used to identify and count discourses in the texts. The ranking of discourses was provided based on the criterion of their semantic-syntactic autonomy. Results. The conducted neural network modeling revealed a high accuracy in predicting the work of the linguistic consciousness functions associated with retrospective self-assessment and anticipatory imagination of the respondents. Another result of this modeling is a partial confirmation of the assumption concerning existence a relationship between the structural parameters of the discursive field (the rank of the discourses and their place in the respective linear sequence) and the leading semantics of the text. Conclusions. A discourse approach to the study of linguistic consciousness, understanding of its structure and functioning features seems to be reasonably appropriate. The implementation of the approach presupposes the need to form a base of linguistic corpora with the inclusion in each text markup of such parameters as: the presence of specific discourses, their ranks, positions in the linear sequence

Discursive conceptualization of LC is not fundamentally new; however, the corresponding empirical studies are still at the early days stage. Among other things, interdiscourse dynamics remains poorly studied, the features of the interconnections of these dynamics with thinking and other mental processes have not been comprehensively investigated. Problems of this kind involve the need to consider at least dozens of different parameters simultaneously. This significantly complicates the use of traditional mathematical and statistical tools in the sense of cumbersome interpretations and generalizations of the findings. An obvious solution to this problem is the use of neural network modeling methodology. A corresponding attempt has been implemented in our work aimed at understanding some of the structural and semantic parameters of interdiscourse interactions.
This article is devoted to the analysis of a series of artificial neural network models (hereinafter -models), built on the corpus of texts obtained in the course of our earlier research (Shymko & Babadzhanova, 2020), in turn directed to clarify the content and structure of the coronavirus quarantine concept as an emerging category of LC. In the current work, we considered forecasting of the leading semantics of respondents' quarantine definitions based on the rank of discourses and their place in the corresponding linear sequence. The same parameters were used to predict the respondents' assessments of how much their psychological well-being and daily life routine changed during the quarantine, as well as to predict the quarantine strategies preferred by them. Thus, we pursued the main goal of this article -to study the validity and reliability of the discourse approach to the phenomenology of LC, to the relevant extent that can be substantiated by the size of the empirical sample and its other qualitative characteristics.

Methods and Techniques of the Research
The primary dataset (Shymko & Babadzhanova, 2020) has been properly modified for the purposes of the current study, the statistical results of which are posted separately (Shymko, 2020b). Below we provide a description of those sampling parameters that have undergone the necessary restructuring and/or which were not detailed in the primary dataset but are essential for this study. So, firstly, the input layer of neural networks in all cases considered here is represented by various parameters that characterize the discursive field, consisting of 9 discourses (Shymko, 2020b: [list of discourses.pdf]). That is, the prediction was carried out based on data of the discourses' presence/ absence, their rank (see below) and place in the corresponding linear sequence. Any others, incl. socio-demographic data were not used in the construction of models for reasons of maintaining the homogeneity of predictors. Secondly, for the identification and counting of discourses in the texts we used the method of 'minimum terminable unit' or T-unit (Hunt, 1977), taking into account the generalizations of the empirical experience of this approach application made by Taboada & Zabala (2008). Thirdly, the ranking of discourses was based on the criterion of their semantic and syntactic autonomy. Namely, the only discourse in the text or one that is commented, explained (causally or in a purposeoriented way), developed (semantically and/or syntactically) by another discourse (discourses) and at the same time itself does not perform such functions -it was evaluated as independent one. All the other discourses were assessed as auxiliary ones.
Example 1 Example 3. «Quarantine is a set of regulatory and restrictive measures that are aimed at preventing the mass spread of an infectious disease and providing for a special regime of economic and other activities, restricting the movement of the population, vehicles, cargo, goods and animals» [Карантин -комплекс регуляторноограничительных мер, которые направлены на недопущение массового распространения инфекционного заболевания и предусматривающие особый режим хозяйственной и иной деятельности, ограничение передвижения населения, транспортных средств, грузов, товаров и животных]. In this example, we found one independent discourse (BUREAUCRATIC RESPONSE) and three auxiliary ones (in the order of their appearance in the text -VIRUS DISSEMINATION, LIFESTYLE CHANGES and CONTACT RESTRICTION).
Taking into account the above, the scale of variables reflecting the presence/absence of discourses in the primary dataset was reformatted from nominal to ordinal one and the data was recoded according to the identified ranks (the absence of discourse in the text was denoted by «zero» rank). Also, additional variables were added to the dataset: (a) «leading discourse» (nominal scale) -to display an independent discourse (if there were two or more such discourses in the text, the first one was accounted); (b) «first discourse» (nominal scale); (c) «second discourse» (nominal scale); (d) third discourse (nominal scale); (e) the number of discourses (metric scale); (f) «sequence of discourses» (nominal scale).
The construction of feedforward artificial neural networks was implemented by using the software IBM SPSS Statistics V 26 (Multilayer Perceptron). Given the small number of texts in the corpus, each model went through ten machine learning cycles. At the same time, each iteration was preceded by a random split of the dataset for crossvalidation into the training and testing parts of the sample (the preset approximate proportion was 70% to 30%, respectively). All models have one hidden layer and a given range of the number of neurons in it: from 1 to 50. This paper discusses the general characteristics of the obtained models and the corresponding predictors importance without delving into the layer's synaptic architecture features. The latter would require a different publication format. However, detailed statistical reports for each iteration are published in the Harvard Dataverse repository and are available for review (Shymko, 2020b).

Results
First of all, those models were generated and compared that predict the texts semantics (leading discourse) based on the ranks of discourses (ordinal scale) and simply by the fact of their presence/ absence (nominal scale). As expected, models with an ordinal scale in the input layer of the neural network have demonstrated a significantly lower percentage of errors (Table 1) with a generalized predictive reliability level of 99.74% and 93.88% at the stage of machine learning and testing, respectively. It is noteworthy that at the 8th iteration, the only model ( Fig. 1) with an ordinal scale in this series was obtained, the predictive reliability of which was 100%, both in the process of machine learning and during testing. Importance indicators of the predictors for this model are shown in Table 2.  Comparative analysis of the predictors importance in different iterations of this and other modeling series indicates the absence of a single (repeatable) hierarchy of variables. Considering the frequency parameters of the discourses' appearance in texts (Shymko & Babadzhanova, 2020: NKMRP4), this indirectly might point to the greater importance of interdiscourse interconnections in comparison with the semantics of the discourses themselves. In other words, to predict the leading discourse in the text, the significance of the structural characteristics of the discursive field prevails over the semantic features of its units.
We find confirmation of the foregoing in the next two series of models, with the help of which we have determined the leading discourse on the basis of nominal distinction in a linear sequence of no more than the two first and no more than the three first discourses (Table 3). Let us recall that the corpus consists of texts with the number of discourses in each from 1 to 5. Table 3. Models summary of the leading discourse prediction (predictors -few first discourses in texts)

Models with the first two discourses
Models with the first three discourses Machine learning 3.16 0.00 Testing 7.86 0.00 As one can see, subject to distinguishing the first three discourses, neural network modeling provides the most reliable prediction of the leading discourse. We additionally verified this conclusion by conducting another 10 iterations in two series of models with an input layer that simultaneously distinguishes the first three discourses and takes into account the presence/absence of discourses (one seriesusing the nominal scale, other one -using the ordinal). In all cases and at all stages level of 100% predictive accuracy were obtained (Shymko, 2020b).
Using a similar architecture of the input layers of neural networks, we continued to predict how much the psychological well-being and daily life routine of the respondents changed during the quarantine span, as well as what were their preferences for respective quarantine strategies. Note that in the latter case, we predicted the corresponding rank, set by the respondents to each of the four strategies they were proposed to evaluate. The results are summarized in Table 4.
In most cases, the quality of models with ordinal scales used for discourses is higher. Some exceptions to the effectiveness of machine learning (DAILY LIFE ROUTINE and SYSTEMIC STRATEGY) may be appeared due to the complication of interdiscourse connections because of the discourses rank parameters consideration. That is, we assume the increase in the number of errors has been observed due to the processing of more complex structures. Nevertheless, statistics on incorrect predictions at the stage of models testing confirm the higher productivity of neural network architectures that distinguish not only interdiscourse interactions, but also the corresponding hierarchical characteristics.

Discussions
Discussion of the results described above requires consideration of the limitations that are objectively inherent to any research. Firstly, it is necessary to pay attention to the quantitative characteristics of the primary sample (Shymko & Babadzhanova, 2020: NKMRP4), which does not allow proper generalizing the conclusions. The same applies to the qualitative aspect of the research empirical base, namely, the presence in the corpus of only Russian-language texts and, as a consequence, the impossibility of making comparisons with texts in other languages of the Slavic group, as well as languages from other groups and branches.
Secondly, certain limitations are directly due to the design of the study itself. Thus, the authors of texts with quarantine definitions were asked to assess changes in their psychological well-being and everyday life routine during quarantine, i.e. the corresponding measurements were carried out within the framework of the phenomenologically associated semantic space. It seems promising to expand such a space by making discourse measurements of LC activity products in other semantic spheres and in variety of others (non-laboratory) situations.
Thirdly, when interpreting the results of this study, it is important to consider the limitations caused by its theoretical (conceptual) focuses and methodological (procedural) aspects. So, we proceeded from the basic ideas about discourse, as of the meaning generated by the dissemination of other meanings (Foucault, 1972). The same mechanism realizes itself at a higher level (in interdiscourse interaction), forming a new structural-semantic «layer» -a discourse formation (Shymko, 2020a). We believe that such dynamics are involved in the organization and functioning of the entire LC. At the same time, we do not exclude alternative views both on the nature of discourses and on the technical methods of their localization and counting in texts. Cross-examination of texts from different theoretical and methodological positions and the subsequent comparative analysis contains a very fruitful perspective from a heuristic point of view.
Given the above constraints, what inferences and extrapolations are possible from the described series of neural network modeling? In our opinion, the useful potential of this study is primarily methodological. So, on the one hand, the effectiveness of the discourse approach to the study of LC phenomena has been empirically tested and confirmed. In particular, this manifested itself in the prediction results of respondents' self-assessments and preferences based on discourse characteristics of their texts. The very fact of the possibility of such a prediction determines the need to formulate several hypotheses regarding the place and role of discursive mechanisms in the work of a thinking as function of LC. We assume that it is advisable to consider thinking not only as operating with information that is organized discursively. Thinking, by itself, functions discursively, i.e. in a procedural sense, thinking is a discursive process. Among other things, such a view would well explain why interdiscourse relationships are more important for predicting the leading discourse than the semantics of the discourses themselves, as discussed above.
One way or another, we convinced that the texts discourseology can be effectively used for predictive assessment of the LC functions and its activity products, in particular: retrospective assessment (changes in psychological well-being and daily life routine) and anticipatory imagination (quarantine strategies ranking). We also believe that the hypothesis about the relationship of the discursive field structural parameters (the discourses ranks and their place in the linier sequence) with the leading semantics of the text has been partially substantiated. However, here it is necessary to distinguish between the discursive field of a text and the discursive field of the corpus. We emphasize that the specified relationship is come out if the composition of both fields is known and accounted for. Thus, the question of the corpus approach to discourseology of LC is actualized. Full verification of this hypothesis presupposes overcoming the above limitations of our research.
On the other hand, this study can be useful in terms of some experience in applying the methodology of artificial neural networks to solve research problems in psycholinguistics on relatively small samples. The need to simultaneously consider a large number of inhomogeneous factors is a typical feature of interdisciplinary research. However, the use of neural networks is not only a matter of the convenience of a holistic view on a patchwork picture of variables. Multifactoriality in psycholinguistics is accompanied by additional complexity in the form of such a phenomenon as the entropy of language (Bentz et al., 2017). All this reduces the sensitivity of traditionally used statistical data processing tools to psycholinguistic patterns. In this respect, artificial neural networks are perhaps the best methodological alternative for now.

Conclusions
A discourse approach to the study of linguistic consciousness, understanding of its structure and functioning features seems to be reasonably appropriate. The implementation of this approach presupposes the need to form a base of linguistic corpora with the inclusion in the markup of each text of such parameters as -the presence of specific discourses, their ranks, positions in the corresponding linear sequence. The conducted neural network modeling on the data of the local corpus (texts of Russian-speaking Ukrainians containing the coronavirus quarantine definitions) reveals a high accuracy in predicting the activity results of some linguistic consciousness functions -retrospective selfassessments and products of the anticipatory imagination. Another result of this modeling is a partial confirmation of the assumption about relationship between the structural parameters of the discursive field (the rank of the discourses and their place in the respective linear sequence) and the leading semantics of the text.
The most important, in our opinion, prospects for further research are associated with overcoming the limitations of current research. Namely, the quantitative and qualitative increase in the corpus base, as well as the diversification of text sources, which would provide the possibility of forming a cross-linguistic discourseology of linguistic consciousness.