自然语言处理

In this module we will have two parts: first, a broad overview of NLP area and our course goals, and second, a text classification task. It is probably the most popular task that you would deal with in real life. It could be news flows classification, sentiment analysis, spam filtering, etc. You will learn how to go from raw texts to predicted classes both with traditional methods (e.g. linear classifiers) and deep learning techniques (e.g. Convolutional Neural Nets).

In this module we will treat texts as sequences of words. You will learn how to predict next words given some previous words. This task is called language modeling and it is used for suggests in search, machine translation, chat-bots, etc. Also you will learn how to predict a sequence of tags for a sequence of words. It could be used to determine part-of-speech tags, named entities or any other tags, e.g. ORIG and DEST in "flights from Moscow to Zurich" query. We will cover methods based on probabilistic graphical models and deep learning.

This module is devoted to a higher abstraction for texts: we will learn vectors that represent meanings. First, we will discuss traditional models of distributional semantics. They are based on a very intuitive idea: "you shall know the word by the company it keeps". Second, we will cover modern tools for word and sentence embeddings, such as word2vec, FastText, StarSpace, etc. Finally, we will discuss how to embed the whole documents with topic models and how these models can be used for search and data exploration.

Nearly any task in NLP can be formulates as a sequence to sequence task: machine translation, summarization, question answering, and many more. In this module we will learn a general encoder-decoder-attention architecture that can be used to solve them. We will cover machine translation in more details and you will see how attention technique resembles word alignment task in traditional pipeline.