Entity lookup in DIET Classifier

Hi all…I’m using rasa for my chat bot developement and i’m using DIET classifier in the pipeline, i have a usecase to add lookup tables foi specific entities and lookup tables work on CRFextractor and i’m using DIET Classifier. My question is, is there a way we can add lookup tables for DIET classifier

I saw some messages about that. Are lookup tables really not working with DIET?


What is not working? I didn’t find opened issue on the github. I looked quickly at the code, the regex seems to do the featurization job also for DIET?

No my question is when using CRF extractor we have to add ner_crf in pipeline to add lookup tables for entity extraction, but now i’m using DIET classifier instead of CRF in my pipeline , so my question is can i use lookup table for entity extraction using DIET classifier in my pipeline or should i add any config on my pipeline

There is nothing to add by default. Just be sure you have RegexFeaturizer in your pipeline

1 Like

@Tanja @akelad @souvikg10 Could you pls confirm if lookup tables can be used with DIET? If yes then how?. Because I have followed everything from the rasa blog on Look Up tables but still it does not recognize any new entity outside of the training data, even if that entity is in the look up table. If i have to add training data for each and every entity, then what is the use of look up table.

any updates on this?

Similar discussion here: Lookup table is not working - #13 by azizullah2017

Adding a lookup table feature might be harder to debug because its influence is not deterministic: the feature should be highly correlated with the named entity class it is designed to indicate, but there is no guarantee that DIET will extract it with 100% consistency. I have found that these lookup table feature help extract some examples but not others. I’m still looking into this to see if I can improve its accuracy. Let me know if anyone has clear evidence that lookup table features are broken or any hints and tips on how to improve their applicability or how to debug DIET. A suggestion for Rasa is to make the pipeline more transparent and debuggable such as showing intermediate internal representations (such as featurized text) and perhaps providing feature-importance plots (easier to do with tree-based classifiers, but still probably possible to do with neural-based classifiers).