Road Map

After semi-regular hiatuses, the S-Space package will be back in regular development. This page serves as a road map of features that are planned for futures versions. Users are encouraged to file issues for features they would like to see in future releases.

3.0.0

• Better support for writing SemanticSpace instances to external-program-friendly formats.
  • Export to Matlab-friendly formats
• Complete re-working of how documents are handled by SemanticSpaces
  • Allow tracking documents by name if the algorithm supports it
  • Allow customized tokenization by the user. This could be just wrapping a List in a Document instance
  • Allow for automatic parsing of documents, rather than requiring CoNLL-formatted input
    • Out of the box support for Stanford or Malt parsers
• Add support for more IAA statistics (e.g., Kappa)
• Add a terrible GUI
  • Seriously, this will be terrible, but at least it will lower the learning curve and potentially make it useful for non-programmers to work with the software
• Better support for compositional models
  • Implementation of the common baseline compositional models:
    • Mitchell and Lapata (2008)
    • Erk and Padó (2008)
    • Baroni and Zamparelli (2010)
  • Some kind of common interface to support easily generating in-context meaning from a pre-built model
• Deprecate (or hide) all of the static-method SVD and MatrixIO code
  • Use the new MatrixFactorization interfaces
  • Optionally start depending on some nice matrix library to do the heavy lifting (might be a 4.0.0 change, ha!)

3.1.0

• Support for automatically tuning the number of dimensions using the method described in Fourtassi and Dupoux (2013) at ACL 2013
• Support for new Deep Learning methods of generating semantic vectors
  • Possible algorithms include
    • Polyglot: Distributed Word Representations for Multilingual NLP
• Support for more compositional models
  • Possibly support tensor operations?