Demo paper accepted to SIGIR 2025!

Written with colleagues at other universities, “MMMORRF: Multimodal Multilingual Modularized Reciprocal Rank Fusion”, was accepted to SIGIR 2025 as a demo paper.

---

This paper is based on work done at the Johns Hopkins University Human Language Technology Center of Excellence’s SCALE 2024 program. We found that a pipeline approach composed of modality-specialized systems out-performed advanced multimodal models on some video retrieval tasks. Additionally, we found that fusing multiple modality-specific rankings derived from such components can be further improved by weighting rankings derived from visual features more heavily when text is expected to be scant based on the query.