Unfortunately, I can’t help with that.

But I’d suggest you also ask for help in the “machine-learning” channel in EEF’s Slack. The Elixir machine-learning community is quite active there.

If the structured text is made of a limited alphabet that can be turned into categorical data, that’s probably the most effective thing to do. What I mean is that something like “200g can”, “1500ml bottle”, etc. can be turned into 3 attributes like “size: 200, unit: g, type: can”, or “size: 1500, unit: ml, type: bottle”, where “size” is numerical, while “unit” and “type” are categorical. Many classification algorithms like decision trees and random forests can use a mix of numerical and categorical data.

If it’s really unstructured text, things will be more difficult. Something like Word2Vec is in principle quite easy to train, but you would need a lot of data in order to train effective embeddings.

In sum, my recommendation would be to parse the structure and treat it as numerical and categorical data if possible at all, as that makes things a lot easier, and enables you to use simpler algorithms that are likely to perform better even with limited data. Only treat it as text if it’s really unstructured text.

Ah I now realize that what you have is an order, represented as a set of items each with its quantity. The challenge is to turn that multiset into a tensor. I assume that the order of the items is irrelevant: if so, sequence models designed to process text (where the order of tokens matters) are probably not the best approach.

If the total number of possible items is known in advance and not too large, you can turn each order into a vector where each element corresponds to a possible item that may be in the order, and its numerical value is the quantity of that item in the order. If, say, you have 1000 different articles, each order would become a vector of 1000 elements. Most values would be 0, meaning that such item is not in the order, while some would be set to the quantity of the corresponding item in the order.

As a simplified example, say that there are only 3 possible items: apple, oranges, and lemons. You then would need a 3-dimensional vector to represent orders, one dimension per possible article. An order that contains 3 apples and nothing else would be represented for example by the vector [3, 0, 0], one that contains 1 orange and 5 lemons would be represented by the vector [0, 1, 5], while one that contains 2 apples and 1 lemon would be represented as the vector [2, 0, 1].

I had quite good results in a similar case, where I was training a model on orders with a total of 20k possible different articles. In my case, to reduce the dimensionality of the input and to capture latent factors, I trained an autoencoder on such 20k-dimensional sparse vector, and used the encoder part to produce smaller but dense vectors to feed to my model. If such path would be promising for your case I can provide more details. Note that you still would need a suitable amount of data for the autoencoder to train meaningful representations.