Summarization with paraphrasing or word limit doesn't work

#2
by keithhon - opened

Text:
The U.K. inflation rate fell for the third month in a row in January to hit 10.1%, below economists’ expectations, but high food and energy prices continued to put the pressure on British households.

Economists polled by Reuters had forecast inflation would drop to 10.3% after the rate fell to 10.5% for December. Inflation has fallen consistently since hitting a 41-year-high of 11.1% in October.

Core CPI, which doesn’t include food, energy, alcohol or tobacco, was 5.3% compared to 5.8% in December, according to the ONS.

I have tried to give the following prompts with the above text.

  1. Summarize this text with paraphrasing:
  2. Summarize this text with 20 words

Both gives me
The U.K. inflation rate fell for the third month in a row in January to hit 10.1%, below economists’ expectations, but high food and energy prices continued to put the pressure on British households.

@philschmid
Do you know why?

What generation arguments did you use?

text = """
The U.K. inflation rate fell for the third month in a row in January to hit 10.1%, below economists’ expectations, but high food and energy prices continued to put the pressure on British households.

Economists polled by Reuters had forecast inflation would drop to 10.3% after the rate fell to 10.5% for December. Inflation has fallen consistently since hitting a 41-year-high of 11.1% in October.

Core CPI, which doesn’t include food, energy, alcohol or tobacco, was 5.3% compared to 5.8% in December, according to the ONS.
"""

batch = tokenizer("Summarize this text with paraphrasing: " + text, return_tensors='pt')

output_tokens = model.generate(**batch, max_new_tokens=250)

print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

Got the same results as you. It feels like the model might not know what summarizing in 20 words means, and you would need to fine-tune it for that.

I see. How about paraphrasing?

Got the same results as you. It feels like the model might not know what summarizing in 20 words means, and you would need to fine-tune it for that.

I thought it's related to the format of the input. But it still behaves the same even if I remove all the newlines from it.

Sign up or log in to comment