Commit 5aefc95
committed
Split bold section headers merged inline with their body: rField 0.753 -> 0.762
4607450 scored 0.40 across the whole poster despite good word capture
(788 words vs 732). Its section headers -- Introduction, Materials,
Results, Discussion, all in _SECTION_KEYWORDS -- are set in Times BOLD at
the SAME 32pt size as the regular-weight body that follows them on the
same line. Block grouping merges them (no font-size change, and
block-level bold is a majority vote the body wins), so the font-size
header heuristic never promotes them and every section collapses into an
unsplit blob that no ground-truth section can match.
Split a block that starts with a known section keyword followed by a
capitalized word into a "## Keyword" header and the body after it. The
capital is the tell that a heading/sentence boundary is there, which
keeps ordinary prose ("Results show that ...", lowercase continuation)
intact. Citation-list keywords (references, bibliography) are excluded,
since a reference list legitimately opens with a capitalized author
surname and a following capital is not a merge signal there.
4607450 title 0.123 -> 0.486, authors+affiliations 0.244 -> 0.809,
Introduction 0.537 -> 0.633, Materials 0.415 -> 0.496; the poster's
overall rField 0.403 -> 0.568. aysaekanger's four infographic sections
also gain (+0.02 to +0.04). Against that, three fields that were already
at or near 1.0 give back a hair from shifted chunk boundaries (42
Introduction -0.010, 4564017 References -0.059, 6724771 -0.007).
Corpus rField 0.753 -> 0.762, w 0.976 -> 0.977, rGlobal unchanged, 20/21
acceptable and the 12/13-end-to-end / 13/13-logic-OK corrector board
held, suite green (260 passed).1 parent d33e018 commit 5aefc95
2 files changed
Lines changed: 1273 additions & 0 deletions
0 commit comments