[Question]: Using flair for creating tags or keywords that represents a given text

[B0rner]

Question

Hey,
I wonder if it is possible to use one of the flair model (for example provided over huggingface) in such a way that it generates keywords from a text that is passed to the data model. Preferably in German.

I tried ner-english-ontonotes-large which works very well on German text for specific entities.

But what I need is model, that returns not specific entities, like the name of the person or the name of the city. I need a model, that returns a couple of keywords, that describes the context of the input text, like input: "I like to have pizza for breakfast." output: "breakfast", "pizza", "favorite meal"` ..which should also work for longer texts.

There is no model in the flair-oniverse, that fits exactly that usecase, right?

But is it possible, to train or fine-tune a flair-model for that? A have a huge set of data, that provides a text and related keywords, which would be ideal for finetuning a model. Is there a way (and a tutorial), how to feed the trainig-data to make the model fit for the usecase?

[fkdosilovic]

... I need a model, that returns a couple of keywords, that describes the context of the input text, like input: "I like to have pizza for breakfast." output: "breakfast", "pizza", "favorite meal"` ..which should also work for longer texts.

What you described is a keyword extraction task. Your best bet is to look at the KeyBERT library.

I wonder if it is possible to use one of the flair model (for example provided over huggingface) in such a way that it generates keywords from a text that is passed to the data model. Preferably in German.

If you want to "generate" your output, maybe look into KeyLLM from the KeyBERT library.

There is no model in the flair-oniverse, that fits exactly that usecase, right?

From the API docs of flair, there doesn't seem to be anything like with KeyBERT.

But is it possible, to train or fine-tune a flair-model for that? A have a huge set of data, that provides a text and related keywords, which would be ideal for finetuning a model. Is there a way (and a tutorial), how to feed the trainig-data to make the model fit for the usecase?

Maybe you can train a multi-label classifier, where inputs are documents and labels are keywords that describe best describe the document.