UDPipe Versions

The UDPipe tool is currently available in two versions:

• UDPipe 1:
  • a standalone C++ application with bindings for Python, Java, C#, Perl
  • models are small and fast (1k words/sec on a single CPU)
  • morphosyntactic performance lacking compared to deep neural networks
• UDPipe 2:
  • a Python prototype only
  • models are of moderate size, requiring more computation
    • 60 words/sec using 1 CPU thread
    • 300 words/sec using 8 CPU threads
    • 2k words/sec on a GPU
  • morphosyntactic performance excellent, with error reduction of ~40-50% compared to UDPipe 1

Both these versions are simultaneously available in our LINDAT REST service.

Note that we are also developing a new version:

• UDPipe 3:
  • again a C++ application with bindings for Python and other languages
  • using libtensorflow dll as a backend
  • in addition to morphosyntactic analysis, other modules are planned (NER, entity linking, ...)