Wenxin Liang

Video question answering involves understanding video content to generate accurate answers to questions. Recent studies have successfully modeled video features and achieved diverse multimodal interaction, yielding impressive outcomes. However, they have overlooked the fact that the video contains richer instances and events beyond the scope of the stated question. Extremely imbalanced alignment of information from both sides leads to significant instability in reasoning. To address this concern, we propose the Video-Context Aligned Transformer (V-CAT), which leverages the context to achieve semantic and content alignment between video and question. Specifically, the video and text are encoded into a shared semantic space initially. We apply contrastive learning to global video token and context token to enhance the semantic alignment. Then, the pooled context feature is utilized to obtain corresponding visual content. Finally, the answer is decoded by integrating the refined video and question features. We evaluate the effectiveness of V-CAT on MSVD-QA and MSRVTT-QA dataset, both achieving state-of-the-art performance. Extended experiments further analyze and demonstrate the effectiveness of each proposed module.

Wenxin Liang, Tingyu Zhang, Han Liu, Feng Zhang. Findings of the Association for Computational Linguistics ACL 2024. 2024.

Influence Maximization (IM) has been widely studied recent several years. Given fixed number of seed users and certain diffusion models, the IM problem aims to select proper seed users in social network such that they can achieve the maximal spread of influence. Most previous work assumes that there are only positive relationships between users, and thus users spread influence positively. However, negative relationships also universally exist in various social networks and are complementary to positive relationships in the information diffusion. In this article, the influence maximization problem is addressed in signed social networks that contain both positive and negative relationships. We propose a novel diffusion model called LT-S and two influence spread functions. The proposed LT-S model extends the classical linear threshold model with opinion formation that incorporates both positive and negative opinions and simulates information diffusion in real-world social networks. The influence spread functions under the LT-S model are neither monotone nor submodular which bring challenges to maximization. RLP algorithm is proposed to tackle the issue, which is improved from R-Greedy algorithm by incorporating two proposed accelerating techniques, the live-edge based and propagation-path based techniques. The results of extensive experiments on public real signed social network datasets demonstrate that our algorithm outperforms the baseline algorithms in terms of both efficiency and effectiveness.