Spacy return numerical value from text numbers

Rasa Open Source
1

Hello all!

I am using Spacy to extract “CARDINAL” entities from user’s input. Currently:

  • If the user writes “3” I get a value of “3”.
  • If the user writes “three” I get a value of “three”.

I need to get the same numerical value (“3”) in the case the user inputs “three” as well. This is necessary for me because I need to integrate with an API that is expecting integers on that slot and I can not do the translation from string (e.g. “twenty five”) to int (e.g. 25) myself. Anyone has any idea how to achieve this?

My config is the following:

- name: SpacyTokenizer
- name: SpacyEntityExtractor
  dimensions: ["CARDINAL"]
- name: SpacyFeaturizer
  pooling: mean
- name: CountVectorsFeaturizer
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 4
- name: DIETClassifier
  epochs: 100

Thanks for the help!

2

I would use Duckling for this. There’s a good blog on this topic here.

3

Hello! Thanks for the help! It says that Rasa NLU uses the REST interface of Duckling. What does that mean exactly? Does it mean that on every input Rasa sends a request to the server running Duckling and gets the required entities? Does Rasa do this automatically (as long as I include this in the config.yml) and I get the annotated entities in the Rasa response? Do I need to perform this communication between Rasa and Duckling myself somehow?

Thanks!

4

Does it mean that on every input Rasa sends a request to the server running Duckling and gets the required entities?

Yes

Does Rasa do this automatically (as long as I include this in the config.yml) and I get the annotated entities in the Rasa response?

And you can use a custom action to review the entity extraction (which you may want to do if you could have multiple entity extractors).

Do I need to perform this communication between Rasa and Duckling myself somehow?

You need to have Duckling running and point to it in the config.yml. Read the duckling section.

1 Like
5

Thanks for the help, really appreciated! One last question, I got Duckling set up and it’s parsing numbers and times. My problem now is that this input: “20221612” is parsed as a number by Duckling, while I want it to be parsed as time. I have also included these sample utterances in my training set:

- coming on [20221216](time)
- leaving on [20220814](time)

but still what I get is 'additional_info': {'value': 20221612, 'type': 'value'}, 'entity': 'number', 'extractor': 'DucklingEntityExtractor'}, {'entity': 'number', 'start': 0, 'end': 8, 'confidence_entity': 0.5148515105247498, 'value': '20221612', 'extractor': 'DIETClassifier'}

Any ideas?

6

Yes, this is part of the reason I suggested you might want to have a custom action where you can resolve the entity to slot conversion.