Анализ неструктурированных данных

2026/2027

ID 1136775

Статус: Маго-лего

Кто читает: Международная лаборатория прикладного сетевого анализа

Где читается: Факультет социальных наук

Охват аудитории: для своего кампуса

Преподаватели: Быстрова Анастасия Владимировна, Шляхов Фёдор Сергеевич

Язык: русский

Кредиты: 6

Контактные часы: 40

Дополнительные материалы в LMS Задать вопрос

Аннотация

This course focuses on applied methods and existing tools for information retrieval: web scrap-ing, data preprocessing, natural language processing. All methods considered in this course require basic knowledge of discrete mathematics and probabilistic theory. For instance, most NLP and IR methods use conditional probability. In this course, we show the implementation of contemporary approaches in existing software packages (preferably in the python frameworks), and demonstrate how these methods can be used for the solution of some real-world problems.

Цель освоения дисциплины

to show the implementation of contemporary approaches in existing software packages (preferably in the python frameworks), and demonstrate how these methods can be used for the solution of some real-world problems.
to introduce the classical methods of natural language processing and data preprocessing (text cleaning and tokenization, vectorization, TF-IDF weighting, statistical language models and classical machine-learning classifiers), and to show how they are implemented in standard Python libraries (NLTK, scikit-learn, spaCy) and applied to real text-processing problems;
to explain the deep-learning methods and algorithms underlying modern NLP (word embeddings, recurrent and convolutional networks, the attention mechanism and the Transformer architecture), and to demonstrate their implementation in Python deep-learning frameworks such as PyTorch and their use for the solution of practical tasks;
to acquaint students with contemporary large language models and the techniques of working with them (pretraining and fine-tuning, parameter-efficient fine-tuning such as LoRA, prompt engineering and retrieval-augmented generation)

Планируемые результаты обучения

be able to criticize constructively and determine existing issues with applied nlp tasks
be able to get necessary data for research and applied projects
be able to perform basic ETL operations with datasets and unstructured data
have an understanding of the basic principles of information retrieval
have the skill to meaningfully develop an appropriate data analysis pipeline
have the skill to work unstructured text data
know advantages of existing natural language processing packages
know the basic principles behind the the existing deep learning approaches

Содержание учебной дисциплины

Text preprocessing. Pandas for text. Regex. TF-IDF.
Syntax parsing. Spacy. Knowledge Graphs.
Embeddings. Clustering. PCA. UMAP
Language Models: n-grams. Marcov chains. Smoothing. Perplexity.
Neural networks for text (FNN, Backprop). Text classification.
RNN. LSTM.
Transformer Architecture. BERT
BERT. Large Language Models.
Vector Databases. MultiAgents. RAG.
Additional topics

Элементы контроля

Homework 1
Students collect a text corpus by parsing a web source (for example, a news website), and then apply the classical NLP methods covered in the first part of the course: text cleaning and preprocessing, vectorization (Bag-of-Words / TF-IDF), and at least one text-classification and one clustering pipeline, with a short interpretation of the results.
Homework 2
Students fine-tune a pretrained language model on a chosen task and dataset using parameter-efficient fine-tuning (for example, LoRA), evaluate the resulting model on an appropriate metric, and briefly analyze the effect of the main training hyperparameters.
Colloquium 1
Colloquium 2

Промежуточная аттестация

2026/2027 2nd module
0.2 * Homework 2 + 0.3 * Colloquium 1 + 0.3 * Colloquium 2 + 0.2 * Homework 1

Список литературы

Авторы

Паринов Андрей Андреевич
Павлова Ирина Анатольевна
Шляхов Фёдор Сергеевич

Программа дисциплины