Donghua Yu

Yuqian Zhang, Donghua Yu, Zhengyuan Lin et al.

Spoken dialogue generation is crucial for applications like podcasts, dynamic commentary, and entertainment content, but poses significant challenges compared to single-utterance text-to-speech (TTS). Key requirements include accurate turn-taking, cross-turn acoustic consistency, and long-form stability, which current models often fail to address due to a lack of dialogue context modeling. To bridge this gap, we present MOSS-TTSD, a spoken dialogue synthesis model designed for expressive, multi-party conversational speech across multiple languages. With enhanced long-context modeling, MOSS-TTSD generates long-form spoken conversations from dialogue scripts with explicit speaker tags, supporting up to 60 minutes of single-pass synthesis, multi-party dialogue with up to 5 speakers, and zero-shot voice cloning from a short reference audio clip. The model supports various mainstream languages, including English and Chinese, and is adapted to several long-form scenarios. Additionally, to address limitations of existing evaluation methods, we propose TTSD-eval, an objective evaluation framework based on forced alignment that measures speaker attribution accuracy and speaker similarity without relying on speaker diarization tools. Both objective and subjective evaluation results show that MOSS-TTSD surpasses strong open-source and proprietary baselines in dialogue synthesis.

Zehong Wang, Qi Li, Donghua Yu et al.

Inspired by the success of contrastive learning (CL) in computer vision and natural language processing, graph contrastive learning (GCL) has been developed to learn discriminative node representations on graph datasets. However, the development of GCL on Heterogeneous Information Networks (HINs) is still in the infant stage. For example, it is unclear how to augment the HINs without substantially altering the underlying semantics, and how to design the contrastive objective to fully capture the rich semantics. Moreover, early investigations demonstrate that CL suffers from sampling bias, whereas conventional debiasing techniques are empirically shown to be inadequate for GCL. How to mitigate the sampling bias for heterogeneous GCL is another important problem. To address the aforementioned challenges, we propose a novel Heterogeneous Graph Contrastive Multi-view Learning (HGCML) model. In particular, we use metapaths as the augmentation to generate multiple subgraphs as multi-views, and propose a contrastive objective to maximize the mutual information between any pairs of metapath-induced views. To alleviate the sampling bias, we further propose a positive sampling strategy to explicitly select positives for each node via jointly considering semantic and structural information preserved on each metapath view. Extensive experiments demonstrate HGCML consistently outperforms state-of-the-art baselines on five real-world benchmark datasets.

Zehong Wang, Qi Li, Donghua Yu

Temporal graph is an abstraction for modeling dynamic systems that consist of evolving interaction elements. In this paper, we aim to solve an important yet neglected problem -- how to learn information from high-order neighbors in temporal graphs? -- to enhance the informativeness and discriminativeness for the learned node representations. We argue that when learning high-order information from temporal graphs, we encounter two challenges, i.e., computational inefficiency and over-smoothing, that cannot be solved by conventional techniques applied on static graphs. To remedy these deficiencies, we propose a temporal propagation-based graph neural network, namely TPGNN. To be specific, the model consists of two distinct components, i.e., propagator and node-wise encoder. The propagator is leveraged to propagate messages from the anchor node to its temporal neighbors within $k$-hop, and then simultaneously update the state of neighborhoods, which enables efficient computation, especially for a deep model. In addition, to prevent over-smoothing, the model compels the messages from $n$-hop neighbors to update the $n$-hop memory vector preserved on the anchor. The node-wise encoder adopts transformer architecture to learn node representations by explicitly learning the importance of memory vectors preserved on the node itself, that is, implicitly modeling the importance of messages from neighbors at different layers, thus mitigating the over-smoothing. Since the encoding process will not query temporal neighbors, we can dramatically save time consumption in inference. Extensive experiments on temporal link prediction and node classification demonstrate the superiority of TPGNN over state-of-the-art baselines in efficiency and robustness.