T5 fine-tuned model: one method ignores min_target_length parameter while one does not

TensorFlow General Discussion
,
1

I created a text summary model by fine-tuning t5 model, and while inferencing I got a one question.

Following two methods were given same input text, min_target_length and max_target_length but two are yielding different result in terms of length.

Here method1 follow my parameter setting. If I set the min_target_length to 64 and max_target_length 256(same setting I used for training) it does generate summary length of 180.

but method2 always generate one sentence summary, no matter how I set the parameter.

Can anyone know me why this is happening?. Thanks

method1

from transformers import pipeline

summarizer = pipeline("summarization", model=model, tokenizer=tokenizer, framework="tf")

summarizer(
    raw_datasets["train"][0]["document"],
    min_length=MIN_TARGET_LENGTH,
    max_length=MAX_TARGET_LENGTH,
)

method2

inputs = ["summarize: " + text]
inputs = tokenizer(inputs, max_length=MAX_INPUT_LENGTH, truncation=True, return_tensors="tf")
output = model.generate(**inputs, num_beams=8, do_sample=True, min_length=64, max_length=256)
decoded_output = tokenizer.batch_decode(output, skip_special_tokens=True)[0]
predicted_summary = nltk.sent_tokenize(decoded_output.strip())[0]
print(predicted_summary)
3

Hi @Seungjun_Lee, In the method2 the reason for getting one sentence summary might be due to this line of code nltk.sent_tokenize(decoded_output.strip())[0]. For example, let us assume the decode_output as

decode_output=' text='sentence1. sentence2. sentence3'

you have applied strip method on decoded output and pass it to nlkt sentence tokenize

nltk.sent_tokenize(decode_output.strip())[0] #output:'sentence1.'

This might be the reason for getting a summary as one sentence. Thank You.