When creating section-level summaries, the ROUGE scores are calculated between each summary sentence and each section as a whole. However, the original DANCER implementation takes the ROUGE scores between every summary sentence and every single sentence in the original text (not section as a whole)
With current DANCER implementation, for each section, there is often only 1 or even no sentence assigned label 1 as part of the extractive section summary.
Solution
Fix Problem 1
Upon fixing Problem 1, each summary sentence is assigned to the top k sections (default 3) instead of just top 1.
Allowing an alternative text-summary alignment/matching: each section is matched with the entire document-level abstractive summary. This is enabled by setting by_section parameter to False for assign_section_level_summaries(); the default value is True)
Context
The implementation of #22 consists of two problem
1
as part of the extractive section summary.Solution
k
sections (default3
) instead of just top 1.by_section
parameter toFalse
forassign_section_level_summaries()
; the default value isTrue
)