Content Moderation in TV Search: Balancing Policy Compliance, Relevance, and User Experience
Adeep Hande, Kishorekumar Sundararajan, Sardar Hamidian, Ferhan Ture
TL;DR
This paper addresses content moderation for TV search where static filtering and metadata errors can misalign results with user intent. The authors propose a retrieval-aware framework with a two-stage pipeline: a meta-heuristic filter that uses time-adaptive lexicon scoring $S(L_j,t)$ and cosine similarity on query, result, and metadata embeddings, plus LLM validators that compute a validation score $V(Q,R_i)$ to refine decisions. A dynamic feedback loop updates $S(L_i,t)$ based on batch-averaged $V(L_i)$ and enables a path toward a distilled model for real-time moderation. Evaluations show the system scales to nearly $10^6$ queries per day, increases true positives over eight weeks, and improves precision, establishing a baseline for context-aware content moderation in TV search.
Abstract
Millions of people rely on search functionality to find and explore content on entertainment platforms. Modern search systems use a combination of candidate generation and ranking approaches, with advanced methods leveraging deep learning and LLM-based techniques to retrieve, generate, and categorize search results. Despite these advancements, search algorithms can still surface inappropriate or irrelevant content due to factors like model unpredictability, metadata errors, or overlooked design flaws. Such issues can misalign with product goals and user expectations, potentially harming user trust and business outcomes. In this work, we introduce an additional monitoring layer using Large Language Models (LLMs) to enhance content moderation. This additional layer flags content if the user did not intend to search for it. This approach serves as a baseline for product quality assurance, with collected feedback used to refine the initial retrieval mechanisms of the search model, ensuring a safer and more reliable user experience.