Please note, this event will not be hosted. This program is not up to date!
Program
Daily Schedule
19th - 23rd of September
| Period | Course |
| 09.00-12.30 | Deep Reinforcement Learning for Dialogue Generation (Heriberto Cuayahuitl, Lincoln University) |
| 14.00-17.30 | Visually Grounded Lexical Semantics (Carina Silberer, Stuttgart University) |
Evenig Lectures
| Sept. 21, 18.00 | From Architect to Nothingness: what happened to Linguistics in Computational Linguistics? (Tibor Kiss, Linguistics Data Science Lab) |
Course Abstracts
Deep Reinforcement Learning for Dialogue Generation (Heriberto Cuayahuitl)
Different subfields within AI have made substantial progress regarding training machines with some abilities to generate sequences of sentences to form dialogues that can be used on text-based or spoken user interfaces. Deep Reinforcement Learning (DRL) is about training systems from (delayed) rewards using neural networks. In this course you will acquire knowledge (and skills) of different methods and algorithms to train DRL-based systems for generating dialogues. While the course will mainly focus on task-oriented dialogues, open-ended dialogue research will also be addressed. Concretely, the following topics will be covered during this self-contained course:
• Introduction to neural nets and reinforcement learning
• Recurrent and Transformer neural nets for dialogue generation
• Value-based deep reinforcement learning for dialogue generation
• Policy-based deep reinforcement learning for dialogue generation
• Frontiers in deep reinforcement learning for dialogue generation
In this course we will alternate between conceptual and practical sessions on each of the topics above. During the practical sessions, you will be provided with example programs and publicly available dialogue data to train example DRL-based dialogue systems. Last, the course will conclude with a discussion on the frontiers of dialogue research.
Visually Grounded Lexical Semantics (Carina Silberer)
How do humans use language to communicate with each other in and about the real world? How can we equip systems that capture this ability to understand and use human language in the physical world? This course addresses these theoretical and practical questions, at whose core lies the fundamental problem of grounding language to perception.
The course thus takes a step beyond computational semantics based on purely language data, and discusses and examines the semantic connection between natural language and visual perception in the world.
We start with a short overview of traditional as well as state-of-the-art text-based semantic representation models (word embeddings), and discuss their properties and limitations with respect to their ability to account for human’s understanding of the meaning and use of language. The course will then give an introduction into computer vision models that are used to extract visual representations of images and the objects they depict. Finally, the main part of the course studies the problem of grounded language learning in the context of connecting words to the visual world, and of integrating concepts and the visual world, respectively. In particular, we will examine (i) human object naming, i.e., the object name(s) humans choose for an individual object, and (ii), visual-linguistic models to learn visually grounded meaning representations of words.
The course will consist of lectures and a practical project for which students are expected to have (some) programming skills, ideally in Python. We recommend that course participants have Python installed, and ideally Anaconda. No prior knowledge of computer vision is required.