EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Temporal Sentence Grounding with Bias-Conflict Synthesis and Collaborative Debiasing

Barrett Dale, Allen Charley, Natsuki Maeda, Amanda Charlotte

International Journal of Computational and Biological Sciences · 01.05.2026 · 3 (2) · 22-33

DOI: 10.66238/ijcbs79 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Computational and Biological Sciences, online erschienen am 01.05.2026. Verfasst von Barrett Dale und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Temporal sentence grounding in videos is a fundamental yet highly challenging task in multimedia understanding, aiming to localize a specific video segment that corresponds to a given natural language query. Despite recent advancements in multimodal architectures, existing models frequently suffer from severe dataset biases. These models tend to memorize spurious correlations between frequently co-occurring visual scenes and linguistic concepts rather than learning genuine cross-modal semantic alignments. To address this critical limitation, this paper introduces a novel framework based on bias-conflict synthesis and collaborative debiasing. We systematically investigate the nature of these shortcut learning behaviors and propose a data synthesis strategy that actively generates bias-conflicting samples. By perturbing both the video and text modalities, we construct counterfactual pairs that break the entrenched prior distributions found in standard training sets. Furthermore, we design a collaborative debiasing training paradigm that employs dual interconnected processing pathways. One pathway is intentionally designed to capture statistical biases, while the other is optimized to overcome these biases by focusing on nuanced, structural multimodal understanding. Extensive theoretical analysis and qualitative discussions are provided to demonstrate how this approach encourages models to rely on true causal relationships. The proposed methodology offers a robust solution for out-of-distribution scenarios, significantly advancing the reliability and generalization capabilities of temporal sentence grounding systems.

Zitieren

APA

Dale, B., Charley, A., Maeda, N., & Charlotte, A. (2026). Temporal Sentence Grounding with Bias-Conflict Synthesis and Collaborative Debiasing. International Journal of Computational and Biological Sciences, 3(2), 22-33. https://doi.org/10.66238/ijcbs79

Vancouver

Dale B, Charley A, Maeda N, Charlotte A. Temporal Sentence Grounding with Bias-Conflict Synthesis and Collaborative Debiasing. International Journal of Computational and Biological Sciences. 2026;3(2):22-33. doi:10.66238/ijcbs79

BibTeX
@article{Dale2026,
  author = {Dale, Barrett and Charley, Allen and Maeda, Natsuki and Charlotte, Amanda},
  title = {Temporal Sentence Grounding with Bias-Conflict Synthesis and Collaborative Debiasing},
  journal = {International Journal of Computational and Biological Sciences},
  year = {2026},
  volume = {3},
  number = {2},
  pages = {22--33},
  doi = {10.66238/ijcbs79}
}
Kostenlos starten →