cogcomp-nlp icon indicating copy to clipboard operation
cogcomp-nlp copied to clipboard

Published 20 hours ago • verified publisher icon CogComp

Retraining models required for fixing #593

Open schen149 opened this issue 6 years ago • 0 comments

TL;DR: We will need to retrain PrepSRL, Quantifier, (maybe) Question Typer

The Fix applied to #593 (currently residing in branch string_join_require_retrain) will affect the behavior of FeatureNGramUtility in edison. All 4 feature generating functions in FeatureNGramUtility calls StringUtils.join() to produce feature names, which currently have trailing dashes. The fix removes the trailing dash, but also changes the feature names.

Here are all packages that will (likely) be affected --

  • PrepSRL https://github.com/CogComp/cogcomp-nlp/blob/7d9dad3fedc16ac59feb278815e27dc195d1367e/prepsrl/src/main/java/edu/illinois/cs/cogcomp/prepsrl/features/WordBigrams.java#L38-L39

  • Quantifier https://github.com/CogComp/cogcomp-nlp/blob/7d9dad3fedc16ac59feb278815e27dc195d1367e/quantifier/src/main/java/edu/illinois/cs/cogcomp/quant/features/WordBigrams.java#L44-L45

  • Question Typer (it looks like just a debug/helper function, not actually used in model? but not sure) https://github.com/CogComp/cogcomp-nlp/blob/7d9dad3fedc16ac59feb278815e27dc195d1367e/question-type/src/main/java/edu/illinois/cs/cogcomp/question_typer/QuestionTyperFeatureExtractorsUtils.java#L191-L200

schen149 avatar Sep 11 '18 18:09 schen149