Which solution work best for this case

[thierry]

Hello all,
I have a question about which solution will fits this case the best:

Case:
I have 3 kinds of text blocks: ingredients, preparation and dosage, and want to classify this types
I have a lot data that are already categorized for training

I hope someone have some papers, GitHub links or even better experience for this case.

best regards !!

[Ekaterina_Dranitsyna]

You can start with this tutorial: Classify structured data with feature columns  |  TensorFlow Core
It demonstrates how to deal with various data types.

[thierry]

I’m sorry I don’t mean data types i want to classify if a text is a ingredient statement preparation statement or dosage statement

[Bhack]

It seems that your problem is just a 3 classes classification task on a single receipit step.
But more in general on this specific topic I suggest to take a look at cooking receipts NER applied to the RecipeDB dataset as it is more interesting:

https://arxiv.org/abs/2004.12184

[8bitmp3]

Maybe experiment with Transformers, which are arguably the most advanced types of architectures for quite a few domains.

Some ideas:

• BERT example: Classify text with BERT  |  Text  |  TensorFlow (uses an “out-of-the-box” IMDB dataset)
• A Keras example from the community: Text classification with Switch Transformer (uses an “out-of-the-box” IMDB dataset)
• (External) Multi-Class Classification With Transformers: Multi-Class Classification With Transformers | Towards Data Science (haven’t tested this but may be useful)

@Ekaterina_Dranitsyna

You can start with this tutorial: Classify structured data with feature columns | TensorFlow Core
It demonstrates how to deal with various data types.

+1 (note that Keras preprocessing layers and/or TF Text API may be more recommended cc @markdaoust)

Pre-processing the dataset and creating a pipeline could be a challenging task especially with a custom dataset. How about also: BERT Preprocessing with TF Text  |  TensorFlow and Text classification with an RNN  |  TensorFlow (if you want to try a less complex model first).