Input data format for multi label classification in Language Studio

matsuo_basho 10

I'm using the language studio to create a multi-label text classification model.

Following the tutorial here.

It appears that the data has to be in a format where each document is its own separate text file and then there is a labels.json file with the particular format described here.

This format is pretty cumbersome for my use-case, where I have a csv with the text and the labels, so just 2 columns. Is there a way to use this type of format (or a variant without creating separate text files for each record) or is the only option the one referenced in the link?

dupammi 6,815 Reputation points Microsoft Vendor

2023-12-22T09:49:15.1133333+00:00

Hi @matsuo_basho ,

Thank you for using Microsoft Q&A forum.

As per your query, the input data format for the multi-label classification in language studio expects a particular format.

As per your requirement of having CSV with the columns "text" and "labels", you will need to follow a custom approach. Try using Python's groupby and apply functionalities, to group the DataFrame by the 'text' column and then aggregate the corresponding 'labels' into a list for each group. Then iterate through the grouped DataFrame to create the desired list of dictionaries.

These built-in functionalities of pandas can be leveraged to generate the labels.json file in a particular format acceptable by the Language studio.

Below are the repro formats that I was able to do at my end

CSV:

JSON output from python:

I hope this helps. Thank you!
matsuo_basho 10 Reputation points

2023-12-22T12:48:11.4733333+00:00

error comment see next one
matsuo_basho 10 Reputation points

2023-12-22T12:52:50.5566667+00:00

@dupammi thanks for the response. What you wrote makes sense. However, I'm not asking how to create the labels json from Python.

I'm asking whether there's an alternative to this data format for a multilabel classification task in Language Services. I do not want to create 10000 small text files and 1 label file. I want to be able to just feed in 1 json file that has the text data with the labels.

If that's not allowed, the lack of this feature is enough to make this modeling with Language Services useless. Maybe there's an alternative way (I already know I can do AutoML multilabel text modeling).
dupammi 6,815 Reputation points Microsoft Vendor

2023-12-22T13:17:34.6+00:00

Hi @matsuo_basho ,

Thank you for your prompt reply.

For multi-label classification tasks in Azure Language Services, the existing format necessitates individual text files and a distinct labels file. Regrettably, it does not accommodate a unified JSON file approach. Nonetheless, your observation highlights viable alternatives, such as Azure AutoML and the Text Analytics API, which can serve as effective substitutes for your specific requirements.

I hope this helps. Thank you!
matsuo_basho 10 Reputation points

2023-12-22T15:06:05.8033333+00:00

@dupammi is the Text Analytics API that you mention something different from modeling with AutoML? Please provide a link if so. Thanks.