Multi-intent spoken language understanding (SLU) can handle multiple intent utterances in real-world scenarios, which has gained increasing research attention. Despite promising results achieved by existing joint models, they (1) perform utterance-level or token-level intent detection, resulting in suboptimal performance due to fixed thresholds or voting mechanisms; (2) incorporate all predicted intents into each slot hidden state and execute parallel slot decoding, which lacks precise intent-slot alignment and overlooks sequential dependencies between slot tokens. In this paper, we propose a new framework to tackle these two issues. For the first issue, we utilize a global pointer with auxiliary tasks to achieve span-based intent detection. For the second issue, we leverage span-based predicted intents for precise intent-slot guidance and introduce rotational position encoding in the interaction module to explicitly model sequential dependencies for precise slot filling. Experimental results on two benchmarks demonstrate the superiority of our framework.

Multi-modal sarcasm detection (MSD) aims to identify sarcasm by analyzing inconsistencies across image-text pairs. Despite promising results achieved, existing methods predominantly focus on model-level improvements, while overlooking data-level challenges. In this paper, we introduce HACL, a Hybrid Adaptive Curriculum Learning framework for MSD. Concretely, we first propose a Hybrid Difficulty Measurer (HDM) that quantifies sarcasm difficulty at both representation and prediction levels, where representation-level difficulty captures cross-modal semantic conflicts, and prediction-level difficulty reflects model learning progress via smoothed loss. Furthermore, an Adaptive Weight Optimizer (AWO) is proposed to dynamically balance these two difficulty metrics, prioritizing implicit sarcasm cues and hard samples. Combined with a competence-based curriculum scheduler, HACL progressively trains models from easy to hard samples. Experiments on two benchmarks demonstrate state-of-the-art performance, with ablation studies validating the necessity of HDM and AWO. Further analyses highlight HACL’s generalizability on multi-modal sentiment analysis (MSA) and its superiority over large vision-language models (LVLMs).

Zhihong Zhu, Kefan Shen, Zhaorun Chen, Yunyan Zhang, Yuyan Chen, Xiaoqi Jiao, Zhongwei Wan, Shaorong Xie, Wei Liu, Xian Wu, Yefeng Zheng. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

In this paper, we pose a new problem of video enhancement transcoding, which converts the compressed dark video into compressed normal-lighting one. Distinct statistics of dark and normal videos result in quite different coding modes, which thus enforces latent constraints on mode conversion during transcoding. Following this idea, we propose a fast mode decision algorithm to speed up computation while maintaining rate-distortion (RD) performance. Experimental results show that our method saves the computational complexity nearly 70%, without noticeable RD loss in comparison with the cascaded decoder-encoder approach.