The extensive audiovisual archives of Radio Televisión Española (RTVE) represent a rich, underutilized resource for production research. A particularly valuable yet challenging collection is the RTVE Film Archive, encompassing news programs from the 1950s to the late 1980s. While undergoing digitalization to improve accessibility, a significant portion of this news collection lacks accompanying sound and suffers from non-informative titles, hindering effective retrieval.

Our central research questions focus on how to effectively explore this news film collection for production needs, evaluate the performance of multimodal information retrieval technology in this context, and determine the feasibility and integration strategy of resulting metadata into the existing archive management system.

Focusing exclusively on non-textual, video-only content, we explore how to effectively represent video content in the absence of accompanying text or metadata by embedding both textual queries and video visual features into a shared vector space. This methodology utilizes the CLIP model to generate embeddings of video frames and natural language queries, constructing a multimodal semantic space where semantic similarity is measured by the proximity of representation vectors. The VTR system pipeline encompasses video analysis, semantic analysis, vector database ingestion, and search with relevance feedback, allowing users to query the film archive using natural language or example images.

The main contributions of this PoC include a demonstration of a functional VTR system for retrieving information from a film archive without relying on existing metadata, an evaluation of the effectiveness of joint semantic space representation for this task, and insights into the potential and challenges of integrating such technologies into existing media archive workflows.