Round4 精度改善レポート

ロードマップ P1–P2(高信頼衝突解消・cell-type strip・fullname 曖昧性・複合句)を engine 0.8.4 で実装した結果です。比較: round3_abbrevround4_abbrev (デプロイなし。判定は Flash×3 固定)。

サマリー

no_match
324
r3 344 → -20 / baseline 663
aligned
1372
r3 1338 → +34 / baseline 1139
wrong
584
r3 642 → -58 / baseline 692
r3→r4 改善/悪化
111 / 34
baseline→r4: 517 / 46

全体ラベル(round3 → round4)

labelround3round4Δ
aligned13381372+34
broader_parent860898+38
partial_or_narrower243248+5
wrong642584-58
ambiguous86-2
source_or_ontology_issue1312-1
no_consensus1014+4
no_match344324-20

abbrev のみ

labelround3round4Δ
aligned913940+27
broader_parent504537+33
partial_or_narrower165169+4
wrong430385-45
ambiguous75-2
source_or_ontology_issue63-3
no_consensus712+5
no_match336317-19

Round4 で入れた変更(engine 0.8.4)

  1. P1 衝突 demotion 強化area TH を primary embed から除外。soft demotion を acronym 衝突+複合句内の裸略語残りに拡大。展開パスは content overlap 必須。
  2. P1 cell-type / marker stripBLA Ppp1r1bCeA-CRFMDTGlutNAc-S D1-SPNs などから分子タグを剥がして ROI 検索。
  3. P1 高信頼衝突ルール追加 — CP/Acb/dSC/pTh/CUN/PCx/OP/ANT/… などコーパス fullname と整合する文献展開を追加。
  4. P2 fullname 曖昧性 — bare paraventricular nucleus / arcuate nucleus を視床下部へ(alias)。
  5. P2 複合句 — 2文字略語の安全なインライン展開(caudal CP)、exact 同点時は name–query Jaccard でタイブレーク(basal nucleus → Meynert)。
  6. やらないこと — HOMBA acronym 削除なし。オントロジーに無い区画への偽 exact なし。デプロイは未実施。

代表例(round3 → round4)

queryround3round4
THwrong
medial division of PHC (area TH)
aligned
thalamus
DLSwrong
(dorso)lateral sulcus
broader_parent
striatum (neostriatum)
CNwrong
cerebral nuclei
aligned
cerebellar deep nuclei
CPwrong
cortical plate of neocortex
aligned
dorsal striatum (caudoputamen complex-CP, STRd)
Acbwrong
anterior lobe of cerebellar cortex
aligned
nucleus accumbens
dSCwrong
dorsal spinocerebellar tract
partial_or_narrower
deep gray layer of superior colliculus
pThwrong
pallidothalamic tract
aligned
ventral thalamus
CeMwrong
central medial nucleus of thalamus
aligned
medial division of central nucleus
BNwrong
basal ganglia (basal nuclei)
aligned
Barrington's nucleus
CUNwrong
cuneus (gyrus)
aligned
cuneiform nucleus
contralateral MECwrong
caudal medial nucleus
aligned
medial (caudal) entorhinal cortex
BLA Ppp1r1bno_match
(no_match)
broader_parent
basolateral nuclear group of amygdala
CeA-CRFno_match
(no_match)
broader_parent
central nucleus of amygdala
MDTGlutno_match
(no_match)
broader_parent
mediodorsal nucleus of thalamus
paraventricular nucleusambiguous
paraventricular nucleus of thalamus
aligned
paraventricular nucleus of hypothalamus
arcuate nucleusambiguous
arcuate nucleus of medulla oblongata
aligned
arcuate nucleus of hypothalamus
caudal CPwrong
tail part of CA1
partial_or_narrower
caudal ventral domain of the striatum
basal nucleuswrong
basal ganglia (basal nuclei)
aligned
basal nucleus of Meynert
VPaligned
ventral pallidum
aligned
ventral pallidum
CA1aligned
CA1 region of hippocampus
aligned
CA1 region of hippocampus

改善例(111)

kindquerybeforeafter
fullnameparaventricular nucleuswrong
paraventricular nucleus of thalamus
aligned
paraventricular nucleus of hypothalamus
abbrevDLSwrong
(dorso)lateral sulcus
broader_parent
striatum (neostriatum)
abbrevCNwrong
cerebral nuclei
aligned
cerebellar deep nuclei
abbrevCPupartial_or_narrower
caudate-putamen cell bridges
aligned
dorsal striatum (caudoputamen complex-CP, STRd)
fullnamecaudate putamenpartial_or_narrower
caudate-putamen cell bridges
aligned
dorsal striatum (caudoputamen complex-CP, STRd)
fullnamearcuate nucleuswrong
arcuate nucleus of medulla oblongata
aligned
arcuate nucleus of hypothalamus
abbrevCPwrong
cortical plate of neocortex
aligned
dorsal striatum (caudoputamen complex-CP, STRd)
abbrevANTwrong
angular nucleus of thalamus
aligned
anterior nuclear complex of thalamus
abbrevTHwrong
medial division of PHC (area TH)
aligned
thalamus
abbrevPuIno_match
(no_match)
broader_parent
pulvinar of thalamus
abbrevspinal interneuronswrong
lamina IX of spinal cord (spinal motor neurons)
broader_parent
spinal cord
abbrevMDTGlutno_match
(no_match)
broader_parent
mediodorsal nucleus of thalamus
abbrevAcbwrong
anterior lobe of cerebellar cortex
aligned
nucleus accumbens
abbrevCeMwrong
central medial nucleus of thalamus
aligned
medial division of central nucleus
abbrevMRwrong
marginal ramus of cingulate sulcus
aligned
median raphe nucleus
abbrevcaudal CPwrong
tail part of CA1
partial_or_narrower
caudal ventral domain of the striatum
fullnamecaudal caudate putamenwrong
caudate-putamen cell bridges
partial_or_narrower
caudal ventral domain of the striatum
abbrevipsi-THsource_or_ontology_issue
medial division of PHC (area TH)
broader_parent
thalamus
abbrevprethalamuswrong
epithalamus
aligned
ventral thalamus
abbrevPPTno_match
(no_match)
aligned
pedunculo(pontine) tegmental nucleus

悪化例(34)

kindquerybeforeafter
fullnamedorsal raphe nucleusaligned
dorsal raphe nucleus
partial_or_narrower
dorsal raphe nucleus, dorsal division
abbrevSp5aligned
spinal trigeminal nucleus
wrong
spinal trigeminal tract
abbrevdorsal raphe nucleusaligned
dorsal raphe nucleus
partial_or_narrower
dorsal raphe nucleus, dorsal division
abbrevMe5aligned
mesencephalic trigeminal nucleus
wrong
mesencephalic trigeminal tract
abbrevNA corealigned
core of nucleus accumbens
partial_or_narrower
lateral subdivision of the core
abbrevOTaligned
olfactory tubercle
wrong
optic tract
abbrevbasal nucleus of the amygdalaaligned
basolateral nucleus (basal nucleus) of amygdala
wrong
basal nucleus of Meynert
fullnameneuropeptide Y promoter-expressing spinal dorsal horn inhibitory interneuronsbroader_parent
dorsal column (horn) of spinal cord
partial_or_narrower
dorsal horn of coccygeal spinal cord
abbrevmatrix thalamusbroader_parent
thalamus
wrong
ventricular (matrix, germinal) zone of spinal cord
abbrevmedial parabrachial nucleusaligned
medial parabrachial nucleus
partial_or_narrower
medial parabrachial nucleus, medial division
fullnamemedial parabrachial nucleusaligned
medial parabrachial nucleus
partial_or_narrower
medial parabrachial nucleus, medial division
abbrevNAc medial shellaligned
medial subdivision of the shell
broader_parent
shell of nucleus accumbens
abbrevsimple lobulealigned
lobule VI (simple lobule)
partial_or_narrower
lobule VIA (simple lobule, rostral part)
abbrevVTA PBPaligned
parabrachial pigmented nucleus
broader_parent
ventral tegmental area
abbrevcervical spinal cord (C6)broader_parent
cervical spinal cord
partial_or_narrower
lamina VI of cervical spinal cord
abbrevcontralateral PNbroader_parent
pons
wrong
pontine nucleus
abbrevdorsal horn (TL)partial_or_narrower
dorsal horn of lumbar spinal cord
wrong
dorsal horn of sacral spinal cord
fullnamedorsal horn (thoracolumbar)partial_or_narrower
dorsal horn of lumbar spinal cord
wrong
dorsal horn of sacral spinal cord
abbrevgemini nucleuswrong
abducens nucleus
no_consensus
paratrigeminal nucleus
fullnameinterfasicular nucleuswrong
anteroventral periventricular nucleus
source_or_ontology_issue
epifascicular nucleus

baseline からの通算

no_match 663 → 324 (-339)、 aligned 1139 → 1372、 wrong 692 → 584。 useful ≈ aligned+broader: 1828 → 2270。