Диссертации, представленные на защиту и подготовленные в НИУ ВШЭ

This dissertation presents a principled framework for explainable machine learning based on Galois connections, specifically leveraging Formal Concept Analysis (FCA) and pattern structures. Focusing on document classification, it proposes inherently explainable models that integrate symbolic reasoning with semantic document representations. There are three core contributions: (1) an FCA-based model that constructs class-specific concept lattices to classify documents and generate interpretable keyword cluster-based explanations, (2) a pattern structure-based aggregate rule classifier over Abstract Meaning Representation (AMR)-based document graphs that identifies stable, semantically meaningful subgraphs and related text fragments as explanations, and (3) a hybrid symbolic-neural model that combines Concept Whitening (CW) and Graph Neural Networks (GNNs) to align the latent dimensions of the GNNs with interpretable graph concepts. The models are evaluated using various metrics for classification accuracy and explanation quality. The results show that Galois connection-based methods can create accurate, inherently interpretable classifiers and provide a formal foundation for explainability in structured data.

The papers collected in Natalia Slioussar’s dissertation are dedicated to the experimental study of inflectional morphology in Russian. Six papers study the relevance of several core concepts from theoretical morphology for language production and processing. These concepts include morphological regularity, productivity, feature markedness, different types of syncretism. As a result of this enterprise, we also learn something new about these concepts, so the main idea of the dissertation is to show how theory and experiment can interact fruitfully. The seventh paper presents a lexical database that contains more than 50 thousand Russian words with more than 1.7 million word forms characterized for more than 70 different parameters. It can be used to facilitate stimulus selection for experimental research. The papers study production and processing of isolated word forms and word forms in a sentence (primarily on the example of number and gender agreement) and in a wider discourse context, using behavioral and neuroimaging methods.