Program for Bot Detection Based on the Analysis of Areas of the Same Intrinsic Dimension in Language Manifolds

In this work, a new research object is presented – the language fractal structure. We hypothesize that the set of vector embeddings of all n-grams of a natural language represents a representative sample of such an object (we use the term Hailonakea to denote the totality of all fractal structures of a language across all n). The study evaluates the intrinsic (true) dimensions of the fractal structures of the Russian and English languages, as well as their regions, with the aim of developing an algorithm to distinguish texts written by humans and bots in the respective languages. For this purpose, we use methods based on (1) topological data analysis, (2) constructing a minimum spanning tree for the considered point cloud (Steele's theorem), and (3) hierarchical graph clustering (Zahn's and Gato- Geva's algorithms). For both languages, the dimensions for all n are found to be non-integer values, which is typical for fractal sets, and are close to 9 for both Russian and English. At the same time, comparing texts of the same language written by humans with those generated by bots reveals a lower dimensionality of bot texts relative to human texts. A prototype detector (binary classifier for humans and bots) was developed during the work, achieving an accuracy of 64% (F1-score: 0.68) for Russian and 79% (F1-score: 0.79) for English. The metrics are calculated for models with mGPT and YaLM architectures.