Skip to content

PromotionDiscussionSingapore

DanFlickinger edited this page Aug 5, 2015 · 1 revision

Suggestion: Organize a tutorial at some *ACl next year

How much documentation is enough?

  • - One measure: when non-DELPH-IN people use the resources - Another: See how often DELPH-IN resources are referenced in publications

Two paths for dissemination

  • - Large projects standardizing on use of particular resource

    • - e.g. Verbmobil, DARPA

    - Grad students choosing 'new' resources for their research

    • - Tutorial would help here - We should also talk to supervisors to send students to the tutorial

Content of tutorial

  • - How to use existing resources? - How to interoperate - how to interpret outputs?

Application developers want modularity

  • - What tasks do our resources help to solve?

The current ACL environment is friendly to semantics

  • - Our dependencies should be at least as good as existing widely-used engines

Who would do the tutorial?

  • - Maybe two would deliver it, but others would help with preparation - Sample applications, improved documentation
    • - relation extraction - sentiment analysis - etc.

Black box packaging of some common configurations of our resources

  • - e.g. multilingual output of MRS in XML

    • - have to acknowledge only partial coverage for some of our languages

    - see pyDelphin as a candidate platform - the term `open-source' can imply high level of technical expertise - we should aim to ease integration with other frameworks such as NLTK - Semeval, LOGON, LTD are examples of attempts at packaging up resources - Note that many other well-known systems are also not simple to install/use - ACE is getting close to turn-key, as well - maybe tune the defaults

Maybe priority is not just ease of use

  • - Benefits will motivate some considerable effort, if can't get it another way

    • - eg Stanford parser, or MALT vs Charniak

    - Need to communicate distinctive services - extra benefit needs to outweigh any higher cost.

Illustrate uses of our MRS outputs or treebanks

  • - DMRS easier to use than MRS - Maybe further packaging: say ACE+pyDelphin - Identify tasks where even partial coverage is benefitial in ensembles

    - SaarGraph: shows alignment of (D)MRS and dependencies

Important to close the gap on unparsed inputs - robustness

  • - less important if using our resources in ensembles

Demonstrate packaging of output APIs, not just standard out - see pyDelphin

Look for extrinsic evaluation settings

Propose SIG on preparing a tutorial, and expect that you will be committed to helping if you attend.

Clone this wiki locally