Abbrev 精度向上レポート

論文コーパス上の略語慣用を優先しつつ HOMBA acronym は維持する方針で、 RCS abbrev 経路を整備した結果です。比較: baselineround3_abbrev (判定は DeepSeek flash × 3-pass、baseline と同条件)。

サマリー

総クエリ
3458
abbrev 2368 / fullname 1090
no_match
344
baseline 663 → -319
aligned
1338
baseline 1139 → +199
label 改善 / 悪化
418 / 25
abbrev のみ 405 / 21

全体ラベル分布

label before% after% Δ
aligned
同一構造・同義語
113932.9%133838.7%+199
broader_parent
正しい上位構造
68919.9%86024.9%+171
partial_or_narrower
一部分・狭すぎる
2437.0%2437.0%0
wrong
異なる構造
69220.0%64218.6%-50
ambiguous
判定困難
80.2%80.2%0
source_or_ontology_issue
入力/オントロジー問題
120.3%130.4%+1
no_consensus
多数意見なし
120.3%100.3%-2
no_match
RCS no_match
66319.2%3449.9%-319

abbrev のみ

label before% after% Δ
aligned
同一構造・同義語
72030.4%91338.6%+193
broader_parent
正しい上位構造
33514.1%50421.3%+169
partial_or_narrower
一部分・狭すぎる
1636.9%1657.0%+2
wrong
異なる構造
47620.1%43018.2%-46
ambiguous
判定困難
70.3%70.3%0
source_or_ontology_issue
入力/オントロジー問題
60.3%60.3%0
no_consensus
多数意見なし
50.2%70.3%+2
no_match
RCS no_match
65627.7%33614.2%-320

システム変更(原理)

方針: HOMBA の unified_ontology_acronym は消さない。 コーパスが示す論文慣用と衝突する場合は、クエリ側の略語展開とスコアリングで文献側を優先する。 オントロジーに存在しない区画へ無理に当てはめない(例: DMS → striatum 親への broader_parent)。

  1. コーパス由来 abbrev ルール追加rcs/homba_abbrev_rules.csv) — no_match / HOMBA acronym 衝突で、fullname 側が妥当なケースから文献展開を抽出。
  2. 文献略語優先の demotion(engine 0.8.x) — フルクエリが略語ルールに一致し、展開側が候補を取れたとき、生略語だけの HOMBA acronym ヒットを抑制。 CA1 のように名前に略語が埋め込まれる語は保護。DMS のような非類似 acronym 衝突は soft demotion。
  3. 略語展開後の alias 再適用 — LHb → lateral habenula → lateral habenular nuclei など。
  4. 展開パスのスコアクエリ — specificity / content 判定を展開語句に対して行い、短い略語の過大ペナルティを回避。
  5. インライン展開の安全弁 — ローマ数字の文中展開禁止; 残トークンが修飾語・側性のみのときだけ文中置換 (DCN interpositus / laminae III-IV を壊さない)。
  6. nucleus/nuclei を non-content 化lateral nuclei accumbens が扁桃体 lateral nucleus に勝つ誤爆を抑制。

代表例(abbrev)

querybeforeafter
VPwrong
ventral posterior visual area (ventral third visual area, V3V)
aligned
ventral pallidum
DMSwrong
dorsal migratory stream
broader_parent
striatum (neostriatum)
PPNwrong
parietopontine fibers
aligned
pedunculo(pontine) tegmental nucleus
CA1aligned
CA1 region of hippocampus
aligned
CA1 region of hippocampus
ARCaligned
arcuate nucleus of hypothalamus
aligned
arcuate nucleus of hypothalamus
IIIwrong
lobule III (central lobule and wing, posterior part)
aligned
oculomotor nucleus
DCN interpositusaligned
interpositus (intermediate) nucleus
aligned
interpositus (intermediate) nucleus
EC IIIbroader_parent
caudal division of entorhinal cortex
broader_parent
caudal division of entorhinal cortex
lateral NAccbroader_parent
nucleus accumbens
broader_parent
nucleus accumbens
dorsolateral PAGaligned
periaqueductal gray substance, dorsolateral division
aligned
periaqueductal gray substance, dorsolateral division
NAc shellaligned
shell of nucleus accumbens
aligned
shell of nucleus accumbens
LHbno_match
(no_match)
aligned
lateral habenular nucleus

改善例(418)

ラベル順位が良くなったクエリ(先頭 25)

kindquerybeforeafter
fullnamenucleus tractus solitariuswrong
parasolitary nucleus
aligned
solitary nucleus
abbrevLHno_match
(no_match)
partial_or_narrower
accessory secretory cells of lateral hypothalamus
abbrevVPwrong
ventral posterior visual area (ventral third visual area, V3V)
aligned
ventral pallidum
abbrevDMSwrong
dorsal migratory stream
broader_parent
striatum (neostriatum)
abbrevdLGNno_match
(no_match)
aligned
lateral geniculate nucleus (dorsal lateral geniculate nucleus)
abbrevTRNno_match
(no_match)
aligned
reticular nucleus of thalamus
abbrevLHbno_match
(no_match)
aligned
lateral habenular nucleus
abbrevDCNno_match
(no_match)
aligned
dorsal cochlear nucleus
abbrevLSno_match
(no_match)
aligned
lateral septal nuclear complex
abbrevLHAno_match
(no_match)
partial_or_narrower
lateral hypothalamic area, tuberal part
abbrevVLPAGno_match
(no_match)
aligned
periaqueductal gray substance, ventrolateral division
abbrevLGNno_match
(no_match)
aligned
lateral geniculate nucleus (dorsal lateral geniculate nucleus)
abbrevPPNwrong
parietopontine fibers
aligned
pedunculo(pontine) tegmental nucleus
abbrevPOmno_match
(no_match)
wrong
periventricular area of posterior thalamus
abbrevvHPCno_match
(no_match)
broader_parent
hippocampus
abbrevSubno_match
(no_match)
aligned
subiculum proper (strictly defined subiculum)
abbrevMGBno_match
(no_match)
aligned
medial geniculate nuclear complex
abbrevPBNno_match
(no_match)
aligned
parabrachial nuclei
abbrevHDBno_match
(no_match)
aligned
horizontal division of nucleus of diagonal band
abbrevMECwrong
caudal medial nucleus
aligned
medial (caudal) entorhinal cortex
abbrevvCA1no_match
(no_match)
aligned
rostral (ventral) division of CA1
abbrevMVNno_match
(no_match)
aligned
medial vestibular nucleus
abbrevNAccpartial_or_narrower
core of nucleus accumbens
aligned
nucleus accumbens
abbrevCPuno_match
(no_match)
partial_or_narrower
caudate-putamen cell bridges
abbrevRVLMno_match
(no_match)
broader_parent
myelencephalon (medulla oblongata)

悪化例(25)

ラベル順位が悪くなったクエリ(先頭 25)。contra-DMS/ipsi-DMS は baseline が migratory stream への誤 aligned で、striatum 親への broader_parent は文献優先の意図的変化。

kindquerybeforeafter
abbrevNAc-S D1-SPNsbroader_parent
nucleus accumbens
wrong
subiculum proper (strictly defined subiculum)
fullnameventral lateral ventral posterior medial thalamic nucleusbroader_parent
ventral posterior nuclear complex of thalamus
wrong
ventral lateral nucleus of thalamus
abbrevCeA-CRFbroader_parent
central nucleus of amygdala
no_match
(no_match)
abbrevNAc-S D2-SPNsbroader_parent
nucleus accumbens
wrong
subiculum proper (strictly defined subiculum)
abbrevBLA INbroader_parent
basolateral nuclear group of amygdala
wrong
intermediate division of lateral nucleus
abbrevBLA Ppp1r1bbroader_parent
basolateral nuclear group of amygdala
no_match
(no_match)
abbrevBLA Rspo2broader_parent
basolateral nuclear group of amygdala
no_match
(no_match)
abbrevBLA.acbroader_parent
basolateral nuclear group of amygdala
wrong
anterior commissure
abbrevBLA.albroader_parent
basolateral nuclear group of amygdala
wrong
ansa lenticularis
abbrevBLA.ambroader_parent
basolateral nuclear group of amygdala
wrong
anteromedial nucleus of thalamus
abbrevCA1-3broader_parent
hippocampal proper (Cornu Ammonis)
partial_or_narrower
CA1 region of hippocampus
abbrevcontra-DMSaligned
dorsal migratory stream
broader_parent
striatum (neostriatum)
fullnamenucleus reticularis tegmenti pontisaligned
reticulotegmental nucleus
wrong
pedunculo(pontine) tegmental nucleus
abbrevdistal CA1broader_parent
hippocampal proper (Cornu Ammonis)
wrong
body part of CA1
abbrevintermediate part of the PVNbroader_parent
paraventricular nucleus of hypothalamus
wrong
interpositus (intermediate) nucleus
abbrevipsi-DMSaligned
dorsal migratory stream
broader_parent
striatum (neostriatum)
abbrevlateral MDTaligned
lateral (multiform) division of MD
broader_parent
mediodorsal nucleus of thalamus
fullnamelocal BLA interneuronsbroader_parent
basolateral nuclear group of amygdala
no_match
(no_match)
abbrevNAc dorsal polebroader_parent
nucleus accumbens
wrong
temporal pole (temporopolar cortex)
abbrevNTS commissural zonealigned
solitary nucleus, commissural division
wrong
subplate (zone) of neocortex
abbrevPVN MC4R+broader_parent
paraventricular nucleus of hypothalamus
no_match
(no_match)
fullnamedescending PVNbroader_parent
paraventricular nucleus of hypothalamus
wrong
lateral fissure, main (caudal) ramus
abbrevseptal CA2broader_parent
hippocampal proper (Cornu Ammonis)
wrong
body part of CA2
abbrevseptofibrial nucleibroader_parent
septal region (septal nuclear complex)
source_or_ontology_issue
basal ganglia (basal nuclei)
abbrevsuperficial CA1broader_parent
hippocampal proper (Cornu Ammonis)
wrong
rostral (ventral) division of CA1

残課題(原理的限界)

成果物