One text classification model, two sets of target classes I am attempting to classify text sequences into two sets of labels - class_1 = [A, B, C, D], class_2 = [X, Y, Z]. The model will be an instance of BertForSequenceClassification from Huggingface's Transformers. For each input sequence, the single model should predict one class from class_1 and one class from class_2, as opposed to training two separate models for each set of classes.
What would be the best way to do this? I am thinking to just have a set of classes to output as:
{
   'A': 0,
   'B': 1,
   'C': 2,
   'D': 3,
   'X': 4,
   'Y': 5,
   'Z': 6
}

The model will then have 7 outputs. I will then deduce the two predicted classes as class_1 = logits[:4].argmax(dim=-1) and class_2 = logits[4:].argmax(dim=-1).

*

*Is this a reasonable approach? Or is there a better way to force a single model to predict one label from each of the 2 different sets of classes?

*If this is a reasonable approach, then what is the best way to handle the loss function? CrossEntropyLoss takes in a model's output logits and corresponding label indices. This wouldn't work in my case. Would it be best to implement my own loss function which is the average of the CrossEntropyLoss between logits[:4] and logits[4:]?

 A: You need two different target variables.
As you point out, if you use a softmax activation or cross-entropy loss on the union of the two label sets, then the information about one label set will be used to predict the other one. This isn't your intention, because a higher (lower) probability of class_1 should not directly cause a lower (higher) probability of class_2.
But the solution is simple. Your proposal in (2) is a perfectly valid solution. A more general form, used in Pareto optimization, lets you prioritize one loss or the other:
$$
\mathcal L = \pi L_y(\hat y, y) + (1- \pi) L_z(\hat z, z)
$$
where $z,y$ are your two labels and $\hat y, \hat z$ are your predictions for the labels and $0 \le \pi \le 1$ is a hyper-parameter that prioritizes learning for $y$ when it is larger and prioritizes learning for $z$ when it is smaller. Your proposal coincides with choosing $\pi = 0.5$.
And for complete generality, $L_z$ and $L_y$ don't even have to be the same loss function. For instance, on some problem, a person might want to predict a categorical feature (like gender) and a continuous feature (like age). In that case, you could use two completely different loss functions.
