Commonsense knowledge extraction and curation

• Type: Seminar
• Teacher: Simon Razniewski (lecturer)
• Credits: 7 ECTS credits
• Registration: Via https://seminars.cs.uni-saarland.de
• Covid: Course participation is possible both offline (preferred), but also online-only.
• Topic Description: Commonsense knowledge (CSK) is critical for building versatile intelligent applications. In delineation from encyclopedic knowledge, which is centered on named entities like Trump, Paris, or FC Barcelona, commonsense is used to refer to properties, traits and relations between general concepts, such as elephants, universities, or painters. Machine-readable collections of CSK are crucial to enable question answering and natural conversation about the world, e.g., by enabling the agent to proactively communicate, identify likely answers, and detect implausible statements and conditions. In this seminar we will study foundational and recent topics around commonsense knowledge extraction and consolidation.

The seminar is a block seminar and will take place on two consecutive days in winter 2021. There will also be two meetings at the beginning of the semester, for which participation is mandatory.

Topics

1. A psychological view on commonsense knowledge
   • Reference 1: McRae, Ken, et al. "Semantic feature production norms for a large set of living and nonliving things." Behavior research methods, 2005
   • Reference 2: Devereux, Barry J., et al. "The Centre for Speech, Language and the Brain (CSLB) concept property norms." Behavior research methods, 2014
2. Modelling commonsense and generics
   • Reference 1: Representations of Commonsense Knowledge, Ernest Davis, Morgan Kaufmann, 1990
   • Reference 2: The Generic Book, Carlson and Pelletier, 1995 (available at our library)
   • For further references on generics, see references in https://arxiv.org/pdf/2005.00660.pdf
3. Time and causality
   • Reference 1: Improving Temporal Relation Extraction with a Globally Acquired Statistical Resource, Ning et al., NAACL 2018
   • Reference 2: CauseNet: Towards a Causality Graph Extracted from the Web, Heindorf et al., CIKM 2020
4. Quantitative and comparative knowledge
   • Reference 1: How Large Are Lions? Inducing Distributions over Quantitative Attributes, Elazar et al., ACL 2019
   • Reference 2: Acquiring Comparative Commonsense Knowledge from the Web. Tandon, Niket, Gerard De Melo, and Gerhard Weikum. AAAI. 2014.
5. Social stereotypes and biases
   • Reference 1: ‘Why do white people have thin lips?’ Google and the perpetuation of stereotypes via auto-complete search forms, Baker and potts, Critical Discourse Studies, 2013
   • Reference 2: Implicit Bias in Crowdsourced Knowledge Graphs. Demartini, Gianluca. Companion Proceedings of The 2019 World Wide Web Conference. 2019.
6. Commonsense explanations
   • Reference 1: Semeval 2020 Task 4 subtask C
     • Individual submissions see ArXiv or Google search for "SemEval 2020 Task 4"
7. Physical commonsense
   • Reference 1: Can a Gorilla Ride a Camel? Learning Semantic Plausibility from Text, Porada et al., EMNLP 2019
   • Reference 2: Do Neural Language Representations Learn Physical Commonsense? Forbes et al., CogSci 2019
8. Noun compounds
   • Reference 1: Olive Oil is Made of Olives, Baby Oil is Made for Babies: Interpreting Noun Compounds Using Paraphrases in a Neural Model, Shwartz and Waterson, NAACL 2018
   • Reference 2: Chapter 3 in Stephen Tratz. 2011.Semantically-enriched parsingfor natural language understanding. University of Southern California.
   • Reference 3: Advanced Semantics for Commonsense Knowledge Extraction, Tuan-Phong Nguyen, Simon Razniewski, Gerhard Weikum, ArXiv, 2020

(own topic suggestions are welcome as well!)

Tentative timeline

• October 30: Participant selection
• November 5, 10am-12: "Introduction to commonsense" lecture, location E1 4 room 024 and virtual via Zoom (choice up to participants) - slides
• November 12, 10am-12: "Seminar survival skills" lecture + 1st deliverable due + topic assignment, location E1 4 room 024 and virtual via Zoom (choice up to participants)
• December 14, 8am: 2nd deliverable due + create metadata in Easychair
• December 14-18: Meetings with advisor
• January 14: 3rd deliverable due + submit bids for reviewing other papers in Easychair
• January 28: 4th deliverable due
• February 11: 5th deliverable due
• March 4: Block seminar presentations

Deliverables

There will be a total of 6 deliverables. To pass the course, all have to be submitted on time. Percentages in brackets denote contribution to final grade.

1. ~2 page writeup on short open questions (5%)
2. Outline of report (5%)
3. Report 1st revision (0%)*
4. Reviews on two other reports (10%)
5. Reports 2nd revision (40%)
6. Presentation (40%)

* 1st revision is not graded, but the prime chance to obtain feedback from advisor and peers.

Material

• Slides 1st meeting
• Slides both meetings
• Questions 1st deliverable
• Report template
• Easychair for reviewing