Multi label text classification

Part 1 (2020)
1

Hello everyone,
I have a data which consists of text column and label column. I have to do a multi label classification for each sample. here is a snapshot of my data .

image
image1455×331 40.8 KB
can Anyone help me out how to input the data for text classifier learner as i have a pre trained language model for this data . I am using fastai v2 . I have seen that there are no tutorials for how to give input for a multi category block text data . without prior knowledge i have done using datablock and here is a snapshot of it and it displayed weird results .
image
image1432×727 86.2 KB
. could anyone help me out for the right format of input to data and what type of loss function should be used…? . when i have printed dls_clas.loss_func it printed FlattenedLoss of BCEWithLogitsLoss() . Is this is the right function or need to change any loss function according to problem…?

1 Like
2

@Anish_sri I think you need to specify label_delim parameter for get_y ColReader

3

@arampacha hello,
can you tell what would be the label_delim = [; ,|] …?

4

in your dataset labels are separated by spaces, so try label_delim=' '

5

@arampacha hello, I have tried by taking label_delim =’ '. I have a doubt regarding the input as my whole motive is to predict multiple labels for a input where my label column has only one label per sample. Is there a possibility that ULMFIT can perform this task. …? As if you see my upper post it is categorizing into different inputs that are not there in the dataset. could you please help me out in solving this problem.

Thanks In Advance,
Anish srivathsav

6

Oh, I thought those were different labels separated by space. In case you have a single label for each item, using any character which is not encountered in your labels as label_delim should work (e.g. label_delim='*')

7

@arampacha Hello,
It dosent’ work even if i have provided other delimiter that have been not there in dataset.

image
image1007×495 36.5 KB
. Is there any possibility that single column values can be one hot encoded and convert the target into multiCategory…? . The values in label column some how have spaces and i just replaced them by ‘_’ .

Thanks,
Anish Srivathsav

8

Ok, you want your system to work if it encounters previously unseen labels, right? Try using
MultiCategoryBlock(add_na=True), I guess it might help

9

@arampacha No, The label which is displaying it as error . all those labels are included in the dataset . Now when i have done this by including add_na = True . it doesn’t help me. Now i got another label was missinfg from dataset.

image
image1122×498 37.3 KB

10

@arampacha Can you help me in deciding this whether it is correct or not…? like i thought of a solution where by taking one hot encodings of all labels and store it in a data frame .

image
image1711×412 25.1 KB
. But i was unable to figure out how to create a datablock for this by using multilabel problem…?

11

@arampacha I have found a similar article related to multi label text classification where they performed it using fast,ai version 1 but I am unable to figure out for version2 . Here is the following code snippet of fastai V1.

image
image886×323 10 KB
. I think if we found a alternative for this will help finding a solution .

Thanks,
Anish srivathsav

12

Well, seems add_na parameter does something different from what I expected.
You can also try passing vocab parameter with all your label to MultiCategoryBlock.

1 Like
13

@arampacha well, I tried that also by passing labels to vocab. That doesn’t help me solving the problem. The same error was getting repeated.

14

Also I think that error is caused by fact, that some of labels end up missing in your train dataset when you do random splitting. How much data do you have? Are there rare labels? You can try to do stratified split

15

@arampacha No I have enough samples in my training data around 5 lakh records with decent number of labels in it. I trained a classifier fit for one cycle , ulmfit got overfitted. I am looking for any alternate approach to do the multi label classification . Thanks for helping me by taking your time to help me out . I am open to any suggestions if you have any to perform a multi label classification where targets are in column with one label per sample.
Thanks,
Anish Srivathsav

16

Can you try datablock.summary(df) and paste the results. It will help to show the pipeline for y and show where things are going wrong.

17

@msivanes Hello , I have a pretrained language model for text classification learner i have Initialized a datablock and ran it . But I got an error while returning the summary.

image
image1292×280 11.1 KB
image
image1001×373 19.2 KB
image
image1459×612 90.8 KB
Here is the snapshot of a sample record
image
image704×358 6.49 KB
. Could you please look into this and suggest the following pipeline accordingly…

Thanks in advance,
Anish

18

@Anish_sri
you may need to call the summary on DataBlock() instance and not on DataLoaders.

dblock = DataBlock(blocks=..., get_x=..., get_y=..,splitter=...)
dblock.summary(source=data)
1 Like
19

@arampacha @msivanes
Hello Manikandan and Arampacha , Thanks for helping me in solving the multi label problem. I have a question regarding the embeddings from the ULMFIT . Is there any way to get the vocabulary tokens of ULMFIT language model embeddings ? If there is a possible way can you please provide any suggested articles or github links to get those embeddings …

Thanks in Advance,
Anish

20

Yes you can.

Is there any way to get the vocabulary tokens of ULMFIT language model embeddings ?

The vocab attribute from the dataloaders instance you create would have the vocabulary.
For example, dls.vocab

This colab showcases how to inspect the tokens from the pretrained model (wiki dataset) and also from the new tokens added from your task specific corpus.

1 Like