What should the format of the input dataset be for Google AutoML Natural Language multi-label text classification? I know that for multi-class classification I need a column of text and another column for labels. The labels column include one label per row.
I have multiple labels for each text and I want to do multi-label classification. I tried having one column per label and one-hot encoding but I got this error message: Max 1000 labels supported. Found 9823 labels.
It was very confusing at first but later I managed to find the format in the documentation, which is a CSV file like:
text1, label1, label2
text2, label2
text3, label3, label2, label1
The parser doesn't understand a table with NULL cells saved as a standard CSV file, which is like:
text1, label1, label2,
text2, label2,,
text3, label3, label2, label1
I had to manually remove extra commas from the CSV file generated by Pandas.
Google AutoML has updated their parser. The following format is fine:
text1, label1, label2, label3,
text1, label1, label2, ,
text1, label1, label2, , ,
At least that worked for me on 27th Jan 2019
One column per label is the way to go. If you have less than 1000 labels, you probably have a mistake in your CSV file, where the parser is getting confused and thinks some of the tokens in the text of the example are labels. Please make sure that your text is correctly escaped with quotes around.
