Overview
Overview
- Midterm exam (13th february)
- Submissions
- Feature-based implementation
- Transformer-based implementation
- Graded individual take-home exam
Literature
- Jurafsky & Martin
- Chapter is in the /literature directory
- https://vu.on.worldcat.org/search/detail/853273078?lang=nl&clusterResults=true&groupVariantRecords=false&queryString=Linguistic%20Fundamentals%20for%20Natural%20Language%20Processing%20100%20Essentials%20from%20Morphology%20and%20Syntax&stickyFacetsChecked=false
- Full access available via VU
- Semantic Role Labeling - Palmer, Martha, Gildea, Xue: https://aclanthology.org/N13-4004/
- There are three 1-hour videos.
- A PDF of the book is in the /literature directory
Resits
- If you pass both assignments, but fail the exam, you can resit the exam.
- If you resit assignment 1 OR assignment 2, submit the missed assignment (capped grade at 6/10) at take home exam deadline.
- If you fail the midterm, you resit on March 6th.
Fail criteria:
- If you fail assignment 1 and assignment 2, you fail the course.
- If you fail the same assignment/exam twice, you fail the course.
Attendence
Mandatory attendance. You can miss up to 2 lectures.
Midterm
- Basics of SRL
- Study https://aclanthology.org/N13-4004/
- Reread and summarize J&M chapter
- Differences between PropBank vs Universal PropBank
- Basics of dependency and constituency parsing
- Essentials of ML
-
Basics of traditional ML (feature engineering and interaction across architectures logreg, svm etc.)
- Basics of transformers/fine-tuning (incl. subword tokenization)
-
SRL
NLP task, closely related with Linguistic Analysis. Integral part of Machine Understanding.
SRL involves detecting events (predicates) and their participants (arguments).
SRL is informed by syntax, verb alternations,
Beth Levin’s verb classes https://www.bol.com/nl/nl/p/english-verb-classes-and-alternations/1001004000643535/
The Case for Case
Case in Dutch is naamval.
Surface-marked cases.
What if cases exist whether or not you see them? What can we do with this? What he defined is that each NP associated with an event/verb could be assigned a semantic role. He give a set of roles to make sense of languaeg and disambiguate.
If you have some way of generalizing over what language can do and what is possible.
Alternations
Collative: (aim) at something. For example, ‘to shoot at X’.
SRL and NLP/Lexical Resources
PropBank
Developed to build a task for NLP. Kind of based on Dowty’s Proto roles. Consistent enough for generalizations but relatively course-grained. Around 20 roles.
Problems with PropBank
- All arguments have notes.
Universal PropBank: FrameNet (Charles Fillmore):