POSID: Target-Task-Unsupervised Person-Attribute Extraction from Heterogeneous Public-Safety Text

KMA Solaiman

cs.CLJul 5, 2025

POSID: Target-Task-Unsupervised Person-Attribute Extraction from Heterogeneous Public-Safety Text

Authors: KMA Solaiman

Organizations: Department of Computer Science, Purdue University, West Lafayette, IN, USA

Abstract

Person descriptions in public-safety text can be compact and regular, or distributed across heterogeneous narratives in which evidence boundaries and person ownership are ambiguous. We present POSID, a target-task-unsupervised framework that decomposes person-attribute extraction into candidate localization, provenance-controlled extraction scope, structured recovery and traversal, and person association, without supervised training or fine-tuning on the target task. We evaluate six localization configurations under two traversal policies across three real-use cohorts and two controlled synthetic diagnostics. On structurally regular descriptions, the strongest configurations reach 1.000/.972 Attr-Only (AO)/Attr-Value (AV) F1 on Compact500. Across the real-use cohorts, every RE-containing configuration benefits from conservative early stopping, while semantic-only retrieval can benefit from continued traversal when localization is less constrained. On the heterogeneous multi-person Challenge14 cohort, the best automatic AO and AV results are .500 and .382; replacing automatic candidates with verified evidence raises the frozen extractor to .522/.410 under early stopping, showing that localization is one of the material bottlenecks. We further introduce provenance-aware scoring for collective and underspecified evidence and a separate person-aware diagnostic for ownership and fragmentation. Together, the results establish a measurable operating regime for inspectable person-attribute extraction and distinguish errors due to localization, traversal, structured recovery, and attribution. Code: https://github.com/ksolaiman/hart-posid.

Figures & tables

Cohort	#Docs	#Person	Role
Base28	28	23	Mixed-source base cohort
Challenge14	14	26	Heterogeneous multi-person challenge cohort
Generated30	30	30	Larger documents created with synthetic context
Compact500	500	500	Compact description of person with synthetic context
Facebook21	21	21	External naturalistic robustness cohort

TABLE I : Evaluation cohorts used in this study.

Fig. 1 : Document structures in the InciText evaluation corpus. Base28 spans compact attribute-dense descriptions, narrative press releases, and semi-structured dispatch/event reports. Challenge14 extends to long incident narratives in which person-attribute evidence can occur at multiple locations. Highlighted regions indicate representative person-description evidence.

Panel of Figure 3 — Fig. 2 : POSID separates evidence localization from structured recovery and person association. Candidate provenance controls extraction scope; the clothing parser supports conservative stopping and continued traversal.

Selector	Downstream Context	Downstream Scope
RE	Matched pattern	wearing : fixed + clothing; fixed pattern: fixed only
Word2Vec	Full sentence	Clothing only
WordNet	Full sentence	Clothing only
Zero-shot NLI	Full sentence	Clothing only

TABLE II : Candidate selectors used by POSID .

Cohort	Best AO	Best AV	Best Configuration(s)	Traversal Policy
Base28	.824	.857	RE+WordNet	early stop
Challenge14	.500	.382	AO: NLI; AV: RE+NLI	AO: continue; AV: early stop
Facebook21	.667	.651	RE, RE+NLI	early stop
Compact500	1.000	.972	RE, NLI, RE+WordNet, RE+NLI	continue
Generated30	1.000	1.000	RE, RE+WordNet, RE+NLI	both identical

TABLE III : Best observed clothing

F_{1}

by cohort. Only Challenge14’s best AO and AV comes from different configurations.

	Base28		Challenge14		Facebook21
Selector	AO	AV	AO	AV	AO	AV
Word2Vec	+.006	+.009	+.100	+.101	-.068	-.053
WordNet	+.012	+.016	+.142	+.119	-.056	-.043
NLI	-.022	-.009	+.050	+.081	-.086	-.044
RE	-.034	-.017	-.107	-.044	-.105	-.063
RE+WordNet	-.037	-.018	-.102	-.042	-.110	-.063

TABLE IV : Continuation effect on clothing

F_{1}

(

\Delta=\mathrm{continue}-\mathrm{early}

). Positive favors continuation; negative favors early stopping. Facebook21 deltas use full-precision primary counts.

Candidate source / extractor / policy	AO TP (Correct items)	AO FP (Unsupported items)	AO FN (Missed items)	AO $F_{1}$	AV $F_{1}$	Person AO	Person AV
NLI / POSID / C	13	17	9	.500	.378	.385	.230
Verified spans / POSID / C	18	30	4	.514	.429	.314	.228
RE+NLI / POSID / E	12	16	10	.480	.382	.160	.103
Verified spans / POSID / E	12	12	10	.522	.410	.261	.148
GPT-5.6 Sol (zero-shot) / –	20	6	2	.833	.735	.792	.613

TABLE V : Challenge14 diagnostic results under identical clothing evaluation. AO TP/FP/FN counts use 22 provenance-adjusted gold clothing-evidence obligations. C: continue; E: early stop.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Attribute group	Person records	Coverage
Gender	23	100.0%
Race	22	95.7%
Height	13	56.5%
Clothing ( wearing )	18	78.3%

Appendix

TABLE VI : Gold annotation coverage for the attribute groups evaluated in Base28. Percentages use the 23 person-level annotation records as the denominator.

Attribute	Rule and normalization
Gender	Match explicit whole-word forms male , man , boy , female , woman , and girl ; canonicalize to male or female .
Race	Accept an explicit race term only in local race–gender adjacency, $R\,G$ or $G\,R$ . Retain the explicit race value.
Height (numeric)	Accept common numeric forms including N feet [N inches] [tall] , N-foot , N ft , and quote notation (e.g., 5’10" ). Preserve the matched expression.
Height (relative)	Accept short or tall only in a local demographic construction of the form (short\|tall) [R] G ; preserve the qualitative value.

Appendix

TABLE VII : Fixed-attribute extraction and normalization rules.

Provenance type	Source-supported evidence	Evaluation obligation
Untagged	Clothing fact is supported for a specific annotated person.	Retain the ordinary instance; person-aware scoring requires the matched person.
shared_identical / shared_subgroup	The same clothing fact is asserted collectively for a tagged group or subgroup.	Count one evidence obligation for that tagged shared fact; repeated stored or predicted realizations do not create artificial cardinality errors.
group_distributed	Several distinct clothing facts are supported for a group, but member-specific ownership is not stated.	Preserve the clothing multiset while making ownership exchangeable within the tagged group.
one_of_group_unspecified	One clothing fact belongs to an unspecified member of a tagged group.	Count one evidence obligation and allow any member of the tagged group to satisfy it.

Appendix

TABLE VIII : Provenance semantics used for clothing evaluation. Special handling applies only to the explicitly tagged field and group.

	Base28				Challenge14
Selector	AO-E	AO-C	AV-E	AV-C	AO-E	AO-C	AV-E	AV-C
RE	.795	.761	.830	.813	.444	.337	.366	.322
Word2Vec	.529	.535	.540	.549	.200	.300	.105	.206
WordNet	.414	.426	.400	.416	.222	.364	.111	.230
NLI	.741	.719	.768	.759	.450	.500	.297	.378
RE+WordNet	.824	.787	.857	.839	.435	.333	.361	.319

Appendix

TABLE IX : Main-cohort clothing

F_{1}

for all six localization configurations and both traversal policies (E: early stop; C: continuation).

Selector	Policy	AO P	AO R	AO $F_{1}$	AV P	AV R	AV $F_{1}$
RE	C	.457	.727	.561	.517	.682	.588
	E	.652	.682	.667	.667	.636	.651
Word2Vec	C	.529	.409	.462	.531	.386	.447
	E	.750	.409	.529	.708	.386	.500
WordNet	C	.357	.227	.278	.400	.227	.290
	E	.625	.227	.333	.625	.227	.333

Appendix

TABLE X : Facebook21 primary clothing extraction under continuation (C) and early stopping (E). The cohort contains 22 AO gold items and 44 AV evaluation units.

Selector	Metric	Continue	Early stop	Change
RE	AO TP	16	15	-1
	AO FP	19	8	-11
	AV TP	30	28	-2
	AV FP	28	14	-14
Word2Vec	AO TP	9	9	0
	AO FP	8	3	-5

Appendix

TABLE XI : Facebook21 false-positive reduction for representative selectors. Counts are from the primary clothing evaluation.

	Generated30 (E=C)		Compact500
Selector	AO	AV	AO-E	AO-C	AV-E	AV-C
RE	1.000	1.000	.908	1.000	.908	.972
Word2Vec	.654	.654	.712	.795	.712	.767
WordNet	.326	.326	.418	.491	.418	.467
NLI	.901	.901	.908	1.000	.908	.972
RE+WordNet	1.000	1.000	.908	1.000	.908	.972

Appendix

TABLE XII : Synthetic-diagnostic clothing

F_{1}

. Generated30 predictions are identical under early stopping and continuation; Compact500 is reported separately for both policies.

Cohort	Attribute	P	R	$F_{1}$
Base28	Gender	1.000	.652	.789
	Race	1.000	.682	.811
	Height	1.000	.692	.818
Challenge14	Gender	1.000	.192	.323
	Race	1.000	.118	.211
	Height	.000	.000	.000

Appendix

TABLE XIII : Fixed-attribute micro precision, recall, and

F_{1}

for the common RE-containing path.

Document Information Extraction

Document Information Extraction turns text, tables, forms, and images into structured, searchable data.

See all 290 papers

Open-Attribute Person Retrieval: Finding People Through Distinctive and Novel Attributes
Aug 2, 2025Minjeong Park, Hongbeen Park, Sangwon Lee +1Information Retrieval Image Retrieval
The Pennsylvania State University · Korea Telecom Research · Korea University +1
Bridging the Pose-Semantic Gap: A Cascade Framework for Text-Based Person Anomaly Search
Apr 25, 2026Zequn Xie, Guijin Luo, Chuxin Wang +4Multimodal IR Anomaly Detection
Zhejiang University · National University of Singapore
LLM-based Schema-Guided Extraction and Validation of Missing-Person Intelligence from Heterogeneous Data Sources
Apr 8, 2026Joshua Castillo, Ravi MukkamalaDocument Information Extraction LLM Information Extraction
Department of Computer Science Old Dominion University Norfolk, Virginia, USA

Aug 2, 2025J/K move · Enter open · S save

Open-Attribute Person Retrieval: Finding People Through Distinctive and Novel Attributes

Minjeong Park, Hongbeen Park, Sangwon Lee, Jinkyu Kim

The Pennsylvania State University · Korea Telecom Research · Korea University+1

Information Retrieval Image Retrieval

Open paper AskPDF arXiv

Person retrieval in surveillance videos often depends on attributes described by witnesses or operators. However, the most useful cues in practice are not always common appearance descriptions (e.g., gender, clothing color), but rare and distinctive attributes that can sharply reduce the search space (e.g., holding a weapon, lying on the ground). Existing text-based person retrieval benchmarks and methods largely focus on identity-centric retrieval with common pedestrian descriptions, leaving such retrieval-critical attributes underexplored. In this paper, we introduce Open-Attribute Person Retrieval (OAPR), a practical retrieval setting that aims to retrieve all pedestrian instances matching a given attribute query, including rare or previously unseen visual concepts, regardless of identity. To support this task, we construct EPAD, an Expanded Pedestrian Attribute Dataset with 267,885 pedestrian images and a unified vocabulary of 65 attributes, including safety-critical actions, assistive devices, and object interactions that are rarely covered in prior benchmarks. We further propose GAP-CLIP, a lightweight CLIP-based framework that learns gated attribute-aware body-part representations for OAPR. Extensive experiments on EPAD demonstrate that GAP-CLIP achieves the strongest top-K retrieval performance on the full attribute space and on out-of-distribution attributes. The code and dataset are available at https://github.com/mlnjeongpark/Open-Attribute-Person-Retrieval.

Abstract

Document Information Extraction

Explore similar work

Open-Attribute Person Retrieval: Finding People Through Distinctive and Novel Attributes

Bridging the Pose-Semantic Gap: A Cascade Framework for Text-Based Person Anomaly Search

LLM-based Schema-Guided Extraction and Validation of Missing-Person Intelligence from Heterogeneous Data Sources