NLU customization for Arabic language

Rasa Open Source
1

Hi everyone,

I am working on NLU customized pipeline for Arabic language. I am trying to test different components and test the performance based on them.

Do you have any advice to be sure that I am working on the right direction all the way? Any suggestions for projects that have used transformers( like bert) on the pipeline for non-English language?

Any other suggestion will be helpful for me.

@koaning

2

SpacyNLP doesn’t seem to have an arabic language model yet, but you can try fastText vectors arabic language models as explained in rasa language support documentation.

1 Like
3

Hi all.

Just to check @Pain have you seen my blogpost on Non-English tools in Rasa?

2 Likes
5

Yes, I have seen it before. It is really helpful.

But, I am looking for suggestions for pipelines that used Arabic language to see the components on it. So, I will have the full understanding where to put the transformer model in the pipeline and I will know what to use with it from available components.

6

Technically, this contains the LaBSE model with is multi-lingual and should contain some notion of Arabic.

pipeline:
- name: WhitespaceTokenizer
- name: CountVectorsFeaturizer
- name: CountVectorsFeaturizer
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 4
- name: LanguageModelFeaturizer
  model_name: bert
  model_weights: rasa/LaBSE
- name: DIETClassifier
epochs: 100

Is that what you mean?

7

Yes, I mean like this. I have tried the following combination, but I think there is an error when using asafaya/bert-base-arabic.

version: "2.0"
language: ar
pipeline:
  - name: WhitespaceTokenizer
  - name: CountVectorsFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: char_wb
    min_ngram: 1
    max_ngram: 4
  - name: LanguageModelFeaturizer
    model_name: "bert"
    model_weights: "rasa/LaBSE"
  - name: DIETClassifier
    epochs: 200
  - name: FallbackClassifier
    threshold: 0.7
  - name: DucklingEntityExtractor
    url: http://localhost:8000
    dimensions:
    - time
    - number
  - name: EntitySynonymMapper

I put the problem on another post, I did that before your reply and maybe it is a different question. In case you want to check, here it is: Problem when using transformer in NLU pipeline

8

Hi Pain Can you please support some help and tell us what Arabic pipeline did you find the best to use finally thanks

9

Hello @Pain, I sow the comments, how things going till now. I faced the same issue.

1 Like
10

Hey Abdo, have you made it through your issue? I’m now developing an Arabic chatbot and I guess I need your advice about that I really appreciate any help you can provide.