Debiased Attention for Temporal Sentence Grounding in Procedure Videos using causal control
Global Media and Social Sciences Research Journal · 10.06.2026 · 7 (2) · 126-138
DOI: 10.71465/gmssrj204 ↗Kurz gefasst
Fachartikel in der Zeitschrift Global Media and Social Sciences Research Journal, online erschienen am 10.06.2026. Verfasst von Riccardo Bellini und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Temporal sentence grounding in procedure videos aims to localize specific video segments corresponding to natural language queries that describe sequential actions. Unlike standard video grounding, procedure videos encompass highly repetitive visual scenes, subtle state changes, and complex temporal dependencies. Current attention-based vision-language models often suffer from severe temporal and visual biases, heavily relying on spurious correlations between frequently co-occurring visual objects and query words rather than genuinely aligning the semantics of the video with the query. In this paper, we propose a novel debiased attention mechanism grounded in causal inference to mitigate these biases in temporal sentence grounding tasks. By modeling the confounding effects of dataset-specific visual priors and linguistic shortcuts, our approach introduces a counterfactual reasoning framework that separates the causal effect of the query from the spurious correlations induced by the environment. We implement a backdoor adjustment strategy integrated directly into the cross-attention layers of a transformer-based architecture. Through extensive empirical evaluation on benchmark procedure video datasets, our proposed method demonstrates significant improvements over existing state-of-the-art models, particularly in out-of-distribution and highly repetitive scenarios. The results highlight the critical importance of causal debiasing in instructional video analysis and provide a robust foundation for future research in trustworthy and generalizable vision-language alignment.
Zitieren
APA
Bellini, R., & Vogel, H. (2026). Debiased Attention for Temporal Sentence Grounding in Procedure Videos using causal control. Global Media and Social Sciences Research Journal, 7(2), 126-138. https://doi.org/10.71465/gmssrj204
Vancouver
Bellini R, Vogel H. Debiased Attention for Temporal Sentence Grounding in Procedure Videos using causal control. Global Media and Social Sciences Research Journal. 2026;7(2):126-138. doi:10.71465/gmssrj204
BibTeX
@article{Bellini2026,
author = {Bellini, Riccardo and Vogel, Hannah},
title = {Debiased Attention for Temporal Sentence Grounding in Procedure Videos using causal control},
journal = {Global Media and Social Sciences Research Journal},
year = {2026},
volume = {7},
number = {2},
pages = {126--138},
doi = {10.71465/gmssrj204}
}
EUVIMED