Graph Deep Learning for NLP (UdS; WiSe 24/25)

Description

Complex pairwise relationships—which are ubiquitous in NLP—are best encoded as graph-structured data: social media networks, relations between entities, and syntactic/semantic parses (to name just a few) are all naturally expressed as graphs. However, constructing effective structural priors for neural models over graphs is considerably more difficult than for “traditional” machine learning media such as text and image data—this is to say that graph data requires highly specialized model architectures. In this seminar, we will delve into active research at the intersection of NLP and graph deep learning, with a particular focus on incorporating graph representations into language models, in order to enable them to engage with graph-structured data such as knowledge graphs and dependency parses.

Prerequisites

This course assumes a solid background in NLP/ML and current language models: you should be familiar with transformer architectures and common methods/terminology in NLP. You should also (at least) know what a graph is. If you don’t, just read the Wikipedia page, and you should be fine: https://en.wikipedia.org/wiki/Graph_(discrete_mathematics).

You do not, however, need any background in graph deep learning: we will discuss current methodologies for modeling graph-stuctured data (i.e. graph neural networks) in class.

Information

Instructor: Michael Sullivan

Time/Location: Thu 16:15 - 17:45; Building C7.2, Room -1.05

Office Hours: Wed 12:00 - 14:00; Building C7.2, Room 1.04 (or on Zoom/Teams by appointment)

Format/Requirements

Each week, we will meet to discuss the assigned reading (see Schedule/Reading List below). All students are expected to ask questions and actively engage in the discussion.

Discussion Leader

At the beginning of the semester, each student will choose a reading for which they will lead the discussion (“Discussion Leader”): if you do not reach out to me to chose a reading, I will randomly assign you a paper. The Discussion Leader will:

While you are not necessarily expected to create slides as Discussion Leader, sometimes visuals can help! Feel free to use a slide deck, if you would like.

The course is divided into topics/themes: if there is another paper that you would like to discuss instead of the assigned reading (that you feel falls under the same topic), come see me during office hours a few weeks before the assigned date. I will update the schedule if I find your paper appropriate for the course and topic.

Evaluation

For students taking the course for 4 credits:

Participation in class discussions: 50%
Discussion Leader: 50%

For students taking the course for 7 credits:

Participation in class discussions: 30%
Discussion Leader: 30%
Term paper: 40%

Schedule/Reading List

Topic Date Reading Background Reading (Optional) Discussion Leader
Introduction Oct. 17 None (logistics and scheduling) - Michael
Introduction Oct. 24 Machine Learning on Graphs: A Model and Comprehensive Taxonomy - Michael
Machine Translation Oct. 31 Document Graph for Neural Machine Translation Semi-Supervised Classification with Graph Convolutional Networks Evan
Machine Translation Nov. 7 A Novel Graph-based Multi-modal Fusion Encoder for Neural Machine Translation - TBD
KG Incorporation Nov. 14 GreaseLM: Graph REASoning Enhanced Language Models Graph Attention Networks Larisa
KG Incorporation Nov. 21 Deep Bidirectional Language-Knowledge Graph Pretraining - Helin
KG Incorporation Nov. 28 Enhancing Dialogue Generation via Dynamic Graph Knowledge Aggregation Weisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks Katja
Relation Extraction Dec. 5 Dialogue Relation Extraction with Document‑Level Heterogeneous Graph Attention Networks - Nellia
Relation Extraction Dec. 12 A Graph Convolutional Network With Multiple Dependency Representations for Relation Extraction - TBD
Structured-Data-to-Text Dec. 19 Improving Encoder by Auxiliary Supervision Tasks for Table-to-Text Generation - TBD
Structured-Data-to-Text Jan. 9 AMR-To-Text Generation with Graph Transformer - TBD
KG Construction Jan. 16 Grapher: Multi-Stage Knowledge Graph Construction using Pretrained Language Models - Shane
KG Construction Jan. 23 Commonsense Knowledge Base Completion with Structural and Semantic Context - Yash
Syntactic/Semantic Parsing Jan. 30 Towards Collaborative Neural-Symbolic Graph Semantic Parsing via Uncertainty - Nhung
Syntactic/Semantic Parsing Feb. 6 Graph-based Dependency Parsing with Graph Neural Networks - Anina

Useful additional reading: Unifying Large Language Models and Knowledge Graphs: A Roadmap

Term Papers

Students taking the course for 7 credits will be expected to write a survey paper on one of the topics discussed in class (or—with my approval—another topic at the intersection of graph deep learning and NLP). Your task is to take a deep dive into the current literature on your chosen topic, and identify 3-4 major sub-areas.

For each sub-area, identify the main challenges and choose 2-3 papers that are representative of that sub-area: you should include relevant readings that we discussed in class, but they do not count towards the 2-3 paper requirement. For each paper, you will summarize and analyze the authors’ methods (experimental design, model architecture, and training procedure) and results: you are essentially expected to act as a “mini Discussion Leader” (minus the “field questions” part) for each of these papers, but in a written format (obviously).

There is no minimum length for the term paper: if you satisfy all of the requirements described above, your paper will be long enough (I’m expecting these to be somewhere in the neighborhood of eight pages). There is also no strict maximum page count. That being said, please limit your paper to a reasonable length: I would really prefer not to have to read fifteen fifty-page papers at the end of the course!

Here are some examples of survey papers, to give you an idea of the expected structure of your report:

Term papers will be due by March 21, 2025, and should be submitted in ACL format.