Commonsense Reasoning for Natural Language Processing Lecture - Dr. Vered Shwartz
This talk covers new advances in NLP research, knowledge in pre-trained language models, commonsense benchmarks, commonsense knowledge resources and integrating commonsense knowledge into neural networks. We also talked about combining text and image input to extract more meaningful representations and reason.
Commonsense knowledge, such as knowing that "bumping into people annoys them" or "rain makes the road slippery", helps humans navigate everyday situations seamlessly. Yet, endowing machines with such human-like commonsense reasoning capabilities has remained an elusive goal of artificial intelligence research for decades. This talk will discuss various challenges related to commonsense reasoning for AI, including how to represent and measure it, as well as incorporate it into downstream tasks.
The talk is based on the ACL 2020 commonsense reasoning tutorial by Vered Shwartz, Maarten Sap, Antoine Bosselut, Dan Roth, and Yejin Choi. https://homes.cs.washington.edu/~msap...
Reference to everything mentioned and covered in the talk:
/ references_from_commonsense_reasoning_for_nlp
/ references_for_the_commonsense_meetup
00:00 Intro
06:43 What is Common Sense?
09:27 Integrating commonsense knowledge into neural networks
12:34 Do pre-trained LMs already capture commonsense knowledge?
15:21 Properties of Concepts (Weir et al., 2020)
21:11 Can we trust knowledge from LMs?
23:43 Zero-shot LM-based Models for commonsense tasks
25:07 Unsupervised Commonsense Question Answering with Self-Talk (Shwartz et al., 2020)
28:32 Knowledge-informed Baselines
34:26 Can you teach LMs symbolic reasoning? (Talmor et al., 2019)
38:43 How do you know that a model is doing commonsense reasoning?
39:28 Step 1: Determine type of reasoning
40:44 Step 2: Choosing a benchmark size
45:32 How to make unlikely answers robust to annotation artifacts?
50:13 COMMONSENSEQA: pivot on knowledge graphs
52:39 Model performance on SOCIAL IQA
53:44 Commonsense benchmarks
58:01 How do you create a commonsense resource?
01:02:38 Extracting commonsense from text
01:04:54 Eliciting commonsense from humans
01:12:02 Incorporating External Knowledge into Neural Models
01:17:42 Limitations of Knowledge Graphs
01:19:22 From Unstructured to Structured Knowledge
01:21:32 Transfer Learning from Language
01:23:10 Visual Commonsense Knowledge Graphs
[Chapters were auto-generated using our proprietary software - contact us if you are interested in access to the software]
Speaker BIO:
Dr. Vered Shwartz is a Postdoctoral researcher at the Allen Institute for AI (AI2) and the University of Washington. Her research interests focus on natural language processing, particularly in lexical and compositional semantics, textual inference, and commonsense reasoning.
Website: https://vered1986.github.io/
----------------------------------------
Find us at:
Newsletter for updates ➜ http://eepurl.com/gJ1t-D
Join our live online lectures ➜ https://www.meetup.com/2d3d-ai/
Subreddit for discussions ➜ / 2d3dai
Discord server for, well, discord ➜ / discord
Blog ➜ https://2d3d.ai
We are the people behind the AI consultancy Abelians ➜ https://abelians.com/