Title A Multimodal Framework and Benchmark for "PREMEPORA-BARONS-01720-PHEVC-WEBDL-BENGALIX": Dataset, Model, and Evaluation
I’m missing context — that term looks like a unique identifier or code rather than a clear topic. I’ll assume you want a publishable abstract + title + short outline for an academic-style paper interpreting "premeporabarons01720phevcwebdlbengalix" as a novel dataset or algorithm name. If you meant something else, tell me.
Abstract We introduce PREMEPORA-BARONS-01720-PHEVC-WEBDL-BENGALIX (hereafter PBB-PWB), a new multimodal dataset and benchmark designed to advance low-resource language understanding, compressed-video processing, and cross-domain web-derived text alignment. PBB-PWB comprises 17,220 annotated video clips encoded with perceptual HEVC variants (PHEVC), paired with crowd-sourced Bengali and code-switched (Bengali–English) transcripts, time-aligned subtitles, and web-derived metadata. We detail dataset curation, compression-aware preprocessing, and three tasks: (1) robust automatic speech recognition for low-bandwidth PHEVC video, (2) multimodal retrieval linking frames and web metadata, and (3) cross-lingual alignment for Bengali–English code-switching. We propose a baseline multimodal architecture combining compression-robust video encoders, wav2vec-style speech encoders fine-tuned on noisy PHEVC audio, and a cross-attention retrieval head. Extensive evaluations show PBB-PWB exposes performance gaps in current state-of-the-art models: relative WER increases of 28–45% under PHEVC artifacts, retrieval mAP drops of 22% for web-noise metadata, and alignment F1 reductions for code-switch segments. We release benchmarks, evaluation scripts, and baseline models to stimulate research in compression-robust multimodal systems for low-resource languages.
