The idea of this work is intriguing. But the settings regarding the sequence lengths in the paper is confusing...
In Sec.2, "During inference, we enumerate all possible spans that are less than the length of L"
In A. 3, "In our experiments, we set the maximum sequence length L to 256"
In A. 4, "For all experiments, we ignore sentence boundaries, and tokenize and split text into sequences with a stride of 16... with the maximum token length of N = 128...The maximum token length for entity spans is set to 30."
Will the setting ignoring sentence boundaries break entity spans (one entity span is split into two sequences)?
If the max token length is 128, a max span length of 256 (L<N) is redundant, right?
What's the difference between "maximum token length for entity spans" and "maximum sequence length L"?
I'd appreciate it if you could address these questions.
The idea of this work is intriguing. But the settings regarding the sequence lengths in the paper is confusing... In Sec.2, "During inference, we enumerate all possible spans that are less than the length of L" In A. 3, "In our experiments, we set the maximum sequence length L to 256" In A. 4, "For all experiments, we ignore sentence boundaries, and tokenize and split text into sequences with a stride of 16... with the maximum token length of N = 128...The maximum token length for entity spans is set to 30."
I'd appreciate it if you could address these questions.