Multi-task Ranking with User Behaviors for Text-video Search

Text-video search has become an important demand in many industrial video sharing platforms, e.g., YouTube, TikTok, and WeChat Channels, thereby attracting increasing research attention. Traditional relevance-based ranking methods for text-video search concentrate on exploiting the semantic relevance between video and query. However, relevance is no longer the principal issue in the ranking stage, because the candidate items retrieved from the matching stage naturally guarantee adequate relevance. Instead, we argue that boosting user satisfaction should be an ultimate goal for ranking and it is promising to excavate cheap and rich user behaviors for model training. To achieve this goal, we propose an effective Multi-Task Ranking pipeline with User Behaviors (MTRUB) for text-video search. Specifically, to exploit the multi-modal data effectively, we put forward a Heterogeneous Multi-modal Fusion Module (HMFM) to fuse the query and video features of different modalities in adaptive ways. Besides that, we design an Independent Multi-modal Input Scheme (IMIS) to alleviate competing task correlation problems in multi-task learning. Experiments on the offline dataset gathered from WeChat Search demonstrate that MTRUB outperforms the baseline by 12.0% in mean gAUC and 13.3% in mean nDCG@10. We also conduct live experiments on a large-scale mobile search engine, i.e., WeChat Search, and MTRUB obtains substantial improvement compared with the traditional relevance-based ranking model.