CSE8803 DLT: Deep Learning for Text Data (2021 Fall)

Table of Contents

Logistics

Learning Objective

This course will introduce state-of-the-art machine learning techniques for mainstay problems in text data analysis, with particular emphasis on deep learning methods that have recently achieved enormous success. Students will learn about trending problems in this field, key methods for solving these problems, and their advantages and disadvantages. The students are also expected to read, present, and discuss research papers, as well as conduct a research oriented course project.

The learning objective is that by the end of this course, the students are able to formulate their text analysis problems at hand, choose appropriate statistical models for the problems, and even come up with innovative solutions for solving open research problems in this field. The course will be useful for students who want to solve practical problems involving text data, and also for those who want to do edge-cutting research in text mining, natural language processing, artificial intelligence, and text-rich interdisciplinary research.

Prerequisites for this course: the students should be familiar with machine learning and have taken a relevant course before (e.g., CX4240, CSE6740, CS4641); (2) the students should be comfortable with reading research papers and giving presentations; (3) the students should have solid programming skills—the course project can be programming demanding.

Course Content

In the task space, we will cover a range of text data analysis tasks, including:

  • Text representation learning
  • Text classification (sentiment analysis, document classification)
  • Sequence labeling (POS tagging, NER, event extraction)
  • Text generation (text summarization, image captioning)
  • Knowledge graph construction and reasoning
  • Question answering and dialogue systems

In the methodology space, we will introduce state-of-the-art deep learning techniques for these problems, including:

  • Deep representation learning (Word2Vec, Transformer, BERT)
  • Sequence labeling models (CRF, LSTM-CRF)
  • Deep transfer/multi-task learning
  • Neural symbolic methods
  • Deep generative models (GPT, VAE, and their variants)

Schedule

Date Topic Presentation Due
08/23/2021 Course Overview   Piazza Signup; Paper Pickup
08/25/2021 Machine Learning Review   Presentation Signup Open Aug 27
08/30/2021 Module 1: Text Representation Learning   Presentation Signup Close
09/01/2021 Word2Vec and Glove P1 and P2 HW1 Out
09/06/2021 No Class (Labor Day)    
09/08/2021 ELMo and BERT P3 and P4  
09/13/2021 Project Guideline and Examples    
09/15/2021 Module 2: Text Classification   HW1 Due
09/20/2021 Embedding-based Classifiers P5 and P6  
09/22/2021 CNN and RNN P7 and P8  
09/27/2021 Attention Models P9 and P10  
09/29/2021 Module 3: Learning with Limited Supervision   HW2 Out
10/04/2021 Transfer Learning and Fine-Tuning P11 and P12  
10/06/2021 Semi-Supervised and Active Learning P13 and P14  
10/11/2021 No class (Fall Break)    
10/13/2021 Weakly Supervised Learning P15 and P16 HW2 Due
10/18/2021 Module 4: Sequence Labeling    
10/20/2021 Neural Sequence Labeling P17 and P18  
10/25/2021 Low-Resource for Sequence Labeling P19 and P20 HW3 Out
10/27/2021 Module 5: Text Generation    
11/01/2021 Sequence-to-sequence models P21 and P22  
11/03/2021 VAE and GPT for Text Generation P23 and P24  
11/08/2021 Multi-modal Text Generation P25 and P26 HW3 Due, project pre-signup open
11/10/2021 Module 6: Knowledge Extraction and Exploitation    
11/15/2021 Knowledge Graph Construction P27 and P28 project pre-signup close
11/17/2021 Knowledge Graph Reasoning P29 and P30 project signup open
11/22/2021 QA and Dialogue with Knowledge Bases P31 and P32 project signup close
11/24/2021 No class    
11/29/2021 Project Presentations    
12/01/2021 Project Presentations    
12/06/2021 Project Presentations    
12/10/2021     Project Report Due

Disclaimer: The instructor reserves the right to modify the planned schedule and grading policy as needed during the course.

Grading

Homework (30%)

  • There will be three assignments. Each one will test your understanding of the taught methods or the presented papers.
  • Late policy: Assignments are due at 11:59PM of the due date. You will be allowed 2 total late days without penalty for the entire semester. Once those days are used, you will be penalized according to the following policy:
    • Homework is worth full credit before the due time.
    • It is worth 75% credit for the next 24 hours.
    • It is worth 50% credit for the second next 24 hours.
    • It is worth zero credit after that.
  • Follow the Georgia Tech Academic Honor Code.

Paper Presentation (20%)

  • We have 32 papers to study, and you will need to pick 1-2 paper(s) from the list to present. The paper list and presentation sign-up sheet is available here (will be open for signup on Aug 27 at 3PM ET).
  • Each presentation is 17 minutes, plus 3 minutes for Q&A. Each presentation can be done by up to three presenters.
  • You need to post your slides on Canvas by 9pm EST the night before your presentation.
  • The presentation will be graded by the instructor and TAs according to the following criteria: quality of slides, presentation clearness, and question addressing. Your presentation should cover at least the following aspects: 1) What is the problem and background? 2) What are the main challenges of the problem? 3) How does the proposed method work? 4) What are the experimental results and observations?
  • If you miss the presentation, unfortunately you will receive zero credit.
  • Useful tips for presentation:

Paper Discussion (10%)

  • After each paper presentation, we will have 10 minutes for free discussion. The discussion will be generally about the strengths/weakness of the paper, what you like/dislike, the practical applications of the techniques, and comparison/connections with other papers we have studied.
  • If you participate in the discussion of one paper and share your thoughts, you will receive 1 discussion point. The total discussion point is 10, which means you need to participate in the discussions for 10 out of the 32 presentations to receive the full score. For each discussion you have participated, please use this form to check-in after you have shared your opinion.

Project (40%)

You need to complete a project on deep learning for text data. Your project needs to be clear about 1) the problem you are attempting to solve; 2) a survey of existing literature for the problem; 3) the technical method you propose in order to solve the problem; 4) the results and conclusion you attain.

  • Each project needs to be completed in a team of 2-4 people. Team members need to clearly claim their contributions in the project report.
  • You will need to do the following:
    • Presentation (20%): group-wise project presentation
    • Final report (20%): a complete and final project report
  • The presentation schedule is available here (will be open for signup Nov 17 at 2pm)
  • Here are some project guidelines and resources you may find useful.

More Resources

Other resources, such as deep learning toolboxes and datasets, will be provided throughout the course.