Daniil Maksymenko

Biography

My primary interest is Natural Language Processing (NLP) since 2019. I started back then with CNNs and RNNs for text classification (sentiment analysis, text source classification), and token classification (NERs task) and then switched to transformer architecture (BERT, XLM-R, Roberta).

My bachelor project was on the verge of NLP and Digital Signal Processing (DSP) as I worked with generative models based on SV2TTS architecture and tuned a model to work with Ukrainian language with voice transfer capabilities. I used techniques such as transfer learning and transliteration to adapt the English-speaking model to Ukrainian without significant tuning from scratch as it needed just 10% of the original dataset to learn new sounds.

My master's degree project was about the controllability of machine translation, where I worked primarily with MarianMT architecture. I checked the ability of this model to adapt to a certain domain with a small chunk of data (fine-tuning efficiency for style and domain control) + created a modified version of MarianMT with semantic search embeddings, where the user can control the target style and domain by passing the prototype-vector of this domain (mean embedding of examples of this domain) and degree of transformation (coefficient [+, +infinity)).

Currently, I entered PhD program in KhNURE at the Computer Science department, Artificial Intelligence course. I plan to work on an efficient adaptation of LLMs for the Ukrainian language using gained insights from my previous research. I research tokenization approaches, fine-tuning efficiency and quality of available models for RAG task with Ukrainian knowledge base.