Natural language processing (NLP) is a significant field of artificial intelligence, and its daily use increases constantly. Current NLP is built around deep neural networks (DNNs), which are complex machine learning systems. However, DNNs are “black boxes”: their internal mechanisms lack human-interpretability. Linguistic analysis typically operates on units like morphemes, words, phrases, or constructions. There is as of yet no proper basis for detecting these in DNNs. I aim to develop the first theoretically grounded framework for the linguistic interpretation of DNN-based NLP systems. I approach this from both a theoretical and an experimental perspective, with the goal of answering three main research questions: 1. What are the criteria for interpreting DNNs linguistically? 2. Which linguistic operations are implemented in current leading NLP applications? 3. How are different language related in multilingual NLP systems?

2022

2025