Skip to content
Snippets Groups Projects
Select Git revision
  • master default protected
  • trie-js-dumping protected
  • hitbox-boosting protected
  • model-metadata protected
  • lm-2-finetuning-whisperggml protected
  • lm-2-finetuning protected
  • lm-2 protected
  • language_model_experiments protected
  • 0.1.24.2 protected
  • 0.1.24.2-rc1 protected
  • 0.1.24.1 protected
  • 0.1.24.1-rc1 protected
  • 0.1.24 protected
  • 0.1.24-rc1 protected
  • 0.1.23.2 protected
  • 0.1.23.1 protected
  • 0.1.23 protected
  • 0.1.22 protected
  • 0.1.21 protected
  • 0.1.20 protected
  • 0.1.19 protected
  • 0.1.18 protected
  • 0.1.17 protected
  • 0.1.16 protected
  • 0.1.15 protected
  • 0.1.14.1 protected
  • 0.1.14 protected
  • 0.1.13 protected
28 results
Search by author
  • Any Author
  • Aleksandras Kostarevas Aleksandras Kostarevas alex
  • Dean Dean dean
  • Tom Folbrecht Tom Folbrecht tom
  • sameer sameer sameer
4 authors
  1. May 26, 2011
  3. May 24, 2011
  5. May 23, 2011
  7. May 21, 2011
  9. May 20, 2011
  11. May 19, 2011
  13. May 18, 2011
  15. May 12, 2011
  17. May 10, 2011
  19. May 09, 2011
  21. Apr 29, 2011
  23. Apr 14, 2011
  25. Apr 06, 2011
  27. Mar 29, 2011
  29. Mar 22, 2011
  31. Mar 17, 2011
  33. Mar 16, 2011
    • Jean Chalard's avatar
      Avoid the removal of high-ranking exactly typed candidates. · d631651b
      Jean Chalard authored 
      It used to be the case that the scoring system turns up the same word
that was entered with a different capitalization, but with a lower
score than some other, more frequent word. To cope with this, there
was code that would order such candidates in the first slot no matter
what. This processing is now useless because fully matching words now
have a huge boost that ensures they will get to the top of the list,
before any non-fully matching word (which means, differing only by
capitalization or accents).

The bug that did happen with this was, if a fully-matching word got
matched by several processing passes, and the (chronologically) later
score affected to this word was weaker, it would result in the
duplicate removal pass removing the stronger score. This in turn would
mess with autocorrect.

In an effort to keep the risk at a minimum for MR1, this change does
not actually remove the useless code, but adds a check in the odd case
to avoid the bad situation. Another change will remove the code for
ICS release.

bug: 4100269

Change-Id: I18c0575332981ffec0e257e26a360995838d521e
      d631651b
  35. Mar 15, 2011
  37. Mar 10, 2011
  39. Mar 09, 2011
  41. Mar 08, 2011