Skip to main content
. Author manuscript; available in PMC: 2024 Sep 30.
Published in final edited form as: Proc Mach Learn Res. 2023 Aug;219:2–30.

Table 3:

Each row represents the Source-Summary alignments computed for metric tuning, whereas the columns denote the alignment method for inference (usage). Each cell represents the instance-level metric correlation to the Human Error Rate, averaged across four metric variants (BARTScore and CTC, Tuned In-Domain and Double Domain). The row-wise max is bolded and column-wise is underlined.

Usage Alignment
R-Gain BS-Gain R-TopK BS-TopK Top Section Entity Chain Tune Avg
Tune Alignment R-Gain .467 .449 .458 .449 .397 .344 .427
BS-Gain .458 .387 .427 .382 .396 .351 .400
R-TopK .449 .440 .442 .446 .408 .387 .428
BS-TopK .460 .411 .435 .407 .416 .387 .419
Top Section .469 .440 .463 .446 .427 .379 .437
Entity Chain .452 .450 .469 .438 .407 .379 .432

Usage Avg .459 .429 .449 .428 .408 .371