Search this site

St. Petersburg Seminars on Natural Language Processing

by ITO University

The goal of seminars is to make a community keeping abreast of modern approaches in the field of natural language processing and computational linguistics. Seminars will propose essences of recent papers, research reports and overviews of relevant topics. Seminars are organised by ITMO University.

Join our Telegram channel with announcements by the link (and a chat).

Where and when?

New location TBA

The seminars are dead?

We are thinking about another format of the seminars. Any updates would be announced at the Telegram channel.

Program

May, 25

15:30 -- 16:15. Natalia Chuprasova (Huawei Technologies). Colorless green ideas sleep furiously: how psycholinguistics works

Psycholinguistics is an interdisciplinary field that combines psychology and linguistics. It investigates human ability to acquire and use language. In this lecture we will examine modern psycholinguistic methods and some of natural language mechanisms which can be useful for NLP-enthusiasts. introduction to psycholinguistics for people with computational background.

16:15 -- 17:00. Amir Bakarov (Higher School of Economics). Linguistic Representativeness of Distributional Semantic Models on the Lexical Level

Distributional Semantics Models (DSMs) like PMI matrices and Word2Vec are one of the most ubiquitous tools in contemporary NLP. However it is still unclear how to optimize such models for specific tasks and how to evaluate them in a general setting. In my work I argue that benefits of intrinsic DSMs evaluation could be questioned since the notion of their "general quality" possibly does not exist. My hypothesis is that DSMs are just being trained to resolve certain tasks (language modeling, sequence labeling, etc) and do not therefore reflect the structure of the lexical level of a human language (in terms of its epistemological foundations). To prove it I am trying to empirically formalize the notion of linguistic representativeness on this level and find out how inner mechanisms of DSMs argue with this notion.

April, 27

12:00 -- 12:45. Sergey Budkov (ITMO University). Never-Ending Learning.

The ways in which computers learn today remain surprisingly narrow when compared to human learning. An alternative paradigm for machine learning called Never-Ending Learning more closely models the diversity, competence and cumulative nature of human learning. In this talk we'll tell in more details about this paradigm and its usage in different applications.

12:45 -- 13:30. Daria Serdyuk (Just AI). Dialog Systems in Natural Habitat.

We'll talk about dialog systems in general: overview, applications, problems. Also, I'll share my story of making an Alisa.Prize winner game.

April, 13

12:00 -- 12:45. Arip Asadulaev (ITMO University). Memory Augmented Neural Networks.

The speaker will discuss a few types of neural network with augmented memory and present some extensions based on adversarial learning. Slides.

12:45 -- 13:30. Alexey Pugachev (Huawei Technologies). Exploring Neural Transducers for End-to-End Speech Recognition (paper overview).

Overview of modern end-to-end speech recognition techniques. We will discuss why their properties are crucial for high-load and embedded systems. Slides.

March, 2 (mark up datathon organised by ITMO Digital Humanities Lab)

11:30 -- 12:15. Evgenia Bogacheva (ITMO University). Coreference Resolution and Some Other Tight Tasks (Part 2).

Coreference resolution is one of the most complex and linguistically oriented tasks in the field of Natural Language Processing. Coreference is a relationship between expressions in text, in which the expressions refer to the same object (situation) of extralinguistic reality. In the first part of the lecture we will consider this phenomenon from the point of view of linguistics and in the second part we will cover the main computational approaches for solving coreference resolution.

12:15 -- 13:00. Denis Teslenko & Vlad Tretyak (Huawei Technologies). Named Entity Recognition in a Nutshell.

Named Entity Recognition (NER) (also known as entity identification, entity extraction, and entity retrieval) is a task finding chunks in unstructured texst and classidying them into predefined categories of entity names (such as persons, organizations, locations, medical codes, time expressions, etc). We will try to understand how to construct NER systems, how to evaluate them, and how deep learning and neural networks help to resolve such task on the example of FLAIR. Slides.

February, 16

11:00 -- 11:45. Mikhail Akulov & Tina Smirnova (ITMO University). Coreference Resolution and Some Other Tight Tasks (Part 1).

In the lecture we will examine two algorithms for coreference resolution. First algorithm is the state-of-the-art end-to-end coreference resolution model. The key idea is to directly consider all spans in a document as potential mentions and learn distributions over possible antecedents for each. The model computes span embeddings that combine context-dependent boundary representations with a head-finding attention mechanism.

The second algorithm is the classical mention-pair model. We will cover different implementations of classifiers and several ways to collect results that are commonly used in mention-pair models. Slides: Mention-Pair Coreference (prezi version), Neural Coreference.

11:45 -- 12:30. Valentin Malykh & Anton Alekseev (PDMI RAS). Aspect-based Rating Prediction Model.

The speakers would present a model which considers text reviews for item recommendations. Slides.

Program (2018)

December, 1

11:00 -- 11:45. Anatoliy Bedrikovskiy & Kirill Korikov (Huawei Technologies). About BERT and Embeddings for Intent Classification.

A story of how we use word embeddings models (ELMO) for intent classification, and an overview of BERT model. Slides: Part 1, Part 2.

11:45 -- 12:30. Anton Belyy (ITMO University). Construction and Evaluation of Complex Paraphrase Detection Systems.

How to construct a simple yet strong baseline for plagiarism detection and a few words on what could co wrong in evaluating this task. Slides: Part 1, Part 2.

December, 13

20:10 - 21:10. Valentin Malykh (VK Research). Highlights of EMNLP 2018.

An overview of the most interesting papers from the largest NLP conference held in November, at Brussels. Slides.

More than 30 people have attended our first meeting!

Organization

Andrey Filchenkov, Aleksandr Bugrovskiy, Ivan Smetannikov,zKirill Korikov, Valentin Malykh, Amir Bakarov.

Google Sites

Report abuse

Google Sites

Report abuse