I have a dataset in the following format:

name, type, signature
Eg1 : A, 2, abc123
Eg2 : A, 2, ab3
Eg3 : A, 2, addc1

If we need to train the following dataset using roberta or any other model how can we do it? Or is there any other way to train this model to detect possible signatures for the name?

  • $\begingroup$ Can you specify your problem/goals? This dataset does not at all look like something I'd use a pretrained transformer architecture like BERT or RoBERTa for. Can you clarify your question by adding a bit of context surrounding the dataset/what the features are/what the target is/whether there is a time-domain (samples go over time or whether they are static). It would help a lot in answering the question. $\endgroup$ Commented May 18, 2023 at 11:32
  • $\begingroup$ It's not time series. Think as if this as a name of product like apple phone/mobile but to detect it it will have iphone in the name and a number like 10,11,12,14, pro, pro max which keeps changing. Now nothing connect the product name to the detection part. and it is all under the cateogry mobile/phones. Now if it is macbook then it'll be under laptop and different models and different year release basically it can also have a 16 digit pin starting with the term mac*************. * might be encrypted value. So there are many variations is product and detection. $\endgroup$
    – jason
    Commented May 18, 2023 at 22:12
  • $\begingroup$ @RobinvanHoornIf I do have a list of product and prodect detection words. How can I predict for new products based on existing product/detection/category? DOes this make sense. For example if it is Microsoft surface then possible match might be something from already trained dataset and category. It may or may not be right but you do have all possible outcomes. $\endgroup$
    – jason
    Commented May 18, 2023 at 22:14
  • $\begingroup$ Reading this question, I can't figure out what your dependent and independent variables are. It might help to edit your question and add a real example of the data you're trying to classify. $\endgroup$
    – Nick ODell
    Commented May 21, 2023 at 23:37

2 Answers 2


Given the nature of the problem, it might not be amenable to machine learning. The structure of the data drives how it can be modeled. The features are "name" (assumed to be a string) and "type" (assumed to be hierarchical categories). The target is called "signature". Typically, signatures are unique. Labeling unique items is commonly called identification (i.e., mapping different occurrences to the same individual instance). That is different than categorization (i.e., finding features common among a discrete group). Machine learning tends to focus on the generalization problem of categorization, not on identification.


To train a model like RoBERTa on this dataset, you can first preprocess the data by tokenizing the names, types, and signatures using a tokenizer specific to the model architecture you want to use. You can then convert the tokenized data into numerical values that can be input into the model.

However, there are several other approaches to training a model to detect possible signatures for product names which may be easier to implement.

One approach is to use rule-based systems, which involve creating a set of rules that identify patterns in the text that are likely to correspond to product names. For example, you could define rules that look for common prefixes or suffixes for product names.

Another approach is to use unsupervised learning techniques, such as clustering or topic modeling, to group similar text together and identify groups that are likely to correspond to product names. However, these approaches may not be as accurate as supervised learning with labeled examples.

Additionally, you could use a combination of these techniques to create a more robust rule-based system for detecting product name signatures.

Keep in mind that the quality of the data and choice of model architecture can impact accuracy, so it is important to experiment with different configurations and that having a large dataset is also important for generalization.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.