Text summarization task

[ramkumarkoppu]

If I want to do text summarization with GPT-2 model using keras_nlp, how do I do it? is there any API available for text summarization?

[jbischof]

@ramkumarkoppu yes, you can do this with any CausalLM model in our library. You can fine tune the base models on your dataset or we already have one Task-level preset available already (link).

There is no special API needed: the general practice is to put a TL;DR token between the original text and summary.

[abheesht17]

Just adding to @jbischof's answer here: if you are looking for summarisation, you can alternatively use BART, an encoder-decoder model (instead of GPT-2). We do have a BART preset available for summarisation (finetuned on the CNN+Daily Mail dataset). You can use it like this:

import keras_nlp

model = keras_nlp.models.BartSeq2SeqLM.from_preset("bart_large_en_cnn")
model.compile("greedy")
model.generate("you text goes here", max_length=100)

Not sure what the quality of the summaries will be, depends on your dataset, and whether it lies in the same domain as CNN+DM. If the generated summaries aren't good, you can always finetune the model.

[mattdangerw]

Yeah +1 to @abheesht17 answer, using the "bart_large_en_cnn" preset is the easiest way to get an article in -> summary out flow today without any training. One thing to keep in mind is that is yet to go in a pypi release, so you will want to follow instructions here for getting the latest changes. Any feedback on this workflow appreciated!

Beyond that, there's a lot of other options you could consider...

• Fine-tuning a seq2seq model (such as BART) on your own dataset.
• Fine-tuning a decoder only model GPT/OPT on a summarization dataset. A common approach is to train on input text that looks like {article_text} TL;DR {summary_text}, which reposes the summarization problem as a single sequence problem.
• Zero-shot prompting, without fine-tuning. Today you could try that out with OPT, but the SOTA would definitely be an instruction fine-tuned LLM. We are actively working on ports on some of these models right now, so stay tuned!