Skip to content

MLFN meeting at ACL 2017.08.05

Normunds Grūzītis edited this page Aug 27, 2017 · 27 revisions

Time:

Saturday, August 5, 2017, ~9:00-17:00

This is the day after the last workshop at ACL.

Location:

Empire Landmark Hotel

1400 Robson Street
Vancouver, BC, Canada V6G 1B9
Tel: 604-687-0511
Fax: 604-687-2801
Robson St. is a busy shopping and restaurant street in the West End neighborhood of Vancouver. According to Google maps, the Empire Landmark is a seven-minute walk from the ACL conference hotel, the Westin Bayshore.

Nathan Schneider's Notes from Discussion period

Application areas:

  • social science, analyzing politics/media, digital humanities?
  • information extraction and inferencing (e.g. tourism domain, finance),knowledge base population, dialogue systems (are there dialogue corpora we should target? ICSI Switchboard, Meetings corpora?)
    • Miriam has done a literature search on uses of FN in dialogue
What could FrameNet do that would be most useful for SRL/frame-semantic parsing and NLP community generally?
  • Sam: Publish the data with FE heads, or better yet, a full dependency-style parse. Finding exact spans is hard.
  • Nathan: With cooperation from someone at Berkeley, may be able to find a student to help with heuristic conversion to some sort of CoNLL format.
  • As a research project, worth looking into conversion between Univer Dependencies and FrameNet GF/PT syntax. Availability of data from various FrameNets. We want as much to be public as possible. Suggestion: Release beta-quality data early to get feedback.
Community infrastructure suggestions:
  • framenet-users mailing list (such as a Google Group)
  • framenet-dev mailing list (for people developing FrameNets)
  • GitHub issue tracker for each FrameNet
Annotation tool:
  • Should other FrameNets move to the snazzy new annotation tool (currently called MFN)? Risk of many competing data formats. MFN ingests Berkeley-style XML. Latvian FrameNet is using WebAnno because it is dependency-based, which is not supported by MFN. Goal should be to make the MFN tool as attractive as possible so FrameNets will move away from legacy tools. Everyone will use MFN for the parallel annotation exercise.
Constructions & other resources: Can we do things now that will support longer-term plans for FrameNet?
  • Improvements to FN quality and consistency discovered by ECG work
    • Quality control check 1: For every pair of frames related by Inheritance or Perspective_on, if both frames have a non-core role of the same name, they should be mapped (?).
    • Quality control check 2: Look for things like Agent FEs that are not connected by frame relations to Intentionally_act.
    • Can you find an undergrad to write & run a bunch of consistency checking scripts, automate some of the improvements?
  • Changes to the FrameNet data model that would support constructional work, or major restructuring. E.g.:
    • changes to kinds of frame relations
    • introduction of frame-evoking grammatical constructions alongside LUs
    • restructure how causative frames and relations are handled
    • treatment of metaphor and relationship to MetaNet
    • better LU coverage (presence in the lexicon and annotation)
    • more and better frames
Upcoming meeting candidates:
  • LSA/SCiL Jan. 2018, Salt Lake City
  • BLS Feb. 2018, Berkeley
  • GURT March 2018, DC
  • LREC May 2018, Japan
  • NAACL Jun. 2018, New Orleans
  • ICCG Jul. 2018, Paris
  • COLING Aug. 2018, Santa Fe
  • EMNLP 2018: Sep./Oct.?

Agenda

12:00-13:30 Lunch (on your own, in one of the many restaurants in the neighborhood)
  • Practicum on the new annotation tool from FrameNet Brasil (and ICSI) (40 minutes)
    • (Now available for the parallel annotation task we decided on at our March meeting in Berkeley.)
  • General discussion of topics from survey (listed below)
14:30-14:45 Coffee break
  • Next steps for the MLFN resource:
    • Display of MLFN data (on the ICSI public website?)
    • Central annotation tool vs. separate annotation tools
    • Coordination again, process for making new frames, what tools are needed? A frame creation wizard?

Proposed Topics, based on responses from survey:

  • Shared Annotation tasks and data extraction from them
    • Multilingual annotation of both frames and constructions
    • What sort of annotation will produce the most useful cross-linguistic information?
    • (Preliminary) results of the parallel annotation study
    • Tools for annotation and alignment of parallel and comparable texts
      • Specifically, Web Tool (FrameNet Brasil) vs. WebAnno 3 (https://webanno.github.io)
  • Determining how general semantic frames are; defining criteria for universal and language-specific frames
    • How to deal with comparable situations with different framing in different languages
    • Differences in frame usages across languages
  • What are the implications for Frame Semantics and Construction Grammar of choosing to represent meaning in frames vs. constructions?
  • What kinds of applications will MLFN have? How can we demonstrate, develop, and encourage others to develop them?
You can’t perform that action at this time.