MétaCan
Menu
Back to cohort

Capturing Narrative Semantics from Captions for Relational Scene Abstraction

2025· preprint· en· W4414317082 on OpenAlex

Why this work is in the frame

A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.

affAt least one author lists a Canadian institution in the pinned OpenAlex snapshot.

Bibliographic record

VenuePreprints.org · 2025
Typepreprint
Languageen
FieldComputer Science
TopicMultimodal Machine Learning Applications
Canadian institutionsSimon Fraser University
Fundersnot available
KeywordsInterpretabilityScene graphClosed captioningSemantics (computer science)GraphScalabilityNatural languageExploitAbstraction

Abstract

fetched live from OpenAlex

Understanding visual scenes as structured graphs of objects and their interactions is central to advancing high-level visual reasoning. Conventional scene graph generation methods rely on dense and carefully annotated supervision, where each subject-predicate-object triplet is coupled with explicit bounding box labels. Such supervision, however, is expensive to obtain and scales poorly to the open world. In contrast, natural image captions provide abundant descriptions of scenes at a fraction of the cost, though they remain weakly aligned and inherently noisy. In this work, we introduce \textbf{LINGGRAPH}, a new framework that transforms captions into an indirect yet powerful supervisory signal for scene graph generation. Unlike prior efforts that reduce supervision to isolated triplets, we exploit the global semantic organization encoded in captions—where entities, modifiers, and actions co-occur in narrative structures—to capture interdependent relationships and commonsense scene dynamics. LINGGRAPH extracts structured linguistic cues from captions, such as nominal groups, adjectival modifiers, and verbal relations, and leverages them to guide the detection and classification of graph components. To mitigate the noise and incompleteness of captions, we devise an iterative refinement process that progressively aligns textual spans with visual regions, discarding irrelevant associations while strengthening meaningful ones. Our study demonstrates that linguistic regularities encoded in captions can effectively substitute fine-grained annotations for training robust relational models. Experiments reveal that integrating both global narrative semantics and local syntactic features yields superior interpretability and accuracy in graph generation, surpassing existing weakly supervised baselines. By disambiguating visually similar entities and ensuring semantic coherence, our approach establishes captions as a scalable and practical form of weak supervision. This work highlights the potential of free-form language as a bridge for structured visual understanding, underscoring its role in unifying vision and language at the relational level.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

Full frame distilled prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.

metaresearch head score (Codex)0.000
metaresearch head score (Gemma)0.000
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesMeta-epidemiology (narrow)
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Observational · Consensus signal: none
GenreCandidate signal: Empirical · Consensus signal: none
Teacher disagreement score0.466
Threshold uncertainty score1.000

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0000.000
Meta-epidemiology (narrow)0.0000.000
Meta-epidemiology (broad)0.0000.000
Bibliometrics0.0000.000
Science and technology studies0.0000.000
Scholarly communication0.0000.000
Open science0.0010.002
Research integrity0.0000.001
Insufficient payload (model declined to judge)0.0000.000

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.140
GPT teacher head0.384
Teacher spread0.244 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it