Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. We present OmniScale, a modular and efficient training framework to accelerate the development of omni-modal LLMs. OmniScale introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. OmniScale also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. Using OmniScale, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

Text classification is a fundamental task in web content mining. Although the existing supervised contrastive learning (SCL) approach combined with pre-trained language models (PLMs) has achieved leading performance in text classification, it lacks fundamental principles. Theoretically motivated by a derived lower bound of mutual information maximization, we propose a dual contrastive learning framework DualCL that satisfies three properties, i.e., parameter-free, augmentation-easy and label-aware. DualCL generates classifier parameters from the PLM and simultaneously uses them for classification and as augmented views of the input text for supervised contrastive learning. Extensive experiments conclusively demonstrate that DualCL excels in learning superior text representations and consistently outperforms baseline models.

Aspect-level sentiment classification (ALSC); a fine-grained task of sentiment analysis holds the promise of machines communicating knowledge of different aspects of a product/service with humans. Specifically, ALSC aims at inferring the sentiment expressed toward a specific aspect term in a user-generated textual content. It is a long-standing problem in sentiment analysis, and its usefulness cannot be underestimated. Recent studies employ the dependency tree and restructure it to model the syntactic relationship among words in text. However, the restructuring process destructs the structural information of the original dependency tree. As a solution, we first construct a syntax graph that preserves the structural information of the dependency tree. We then propose a reliable syntax-based neural network model that performs a thorough search on the syntax graph to effectively find the relevant contextual information with respect to the aspect term for the sentence encoding. Noting that dependency trees parsed from existing dependency parsers may contain incorrect syntactic dependencies due to grammatical errors in a sentence, we adopt a convolutional layer that takes into account the relations among features of words in a local neighborhood to mitigate the issues brought by incorrect syntactic dependencies. Our results on benchmark datasets demonstrate that our model outperforms the previous methods and achieves state-of-the-art results for the ALSC task.

Effective regularization techniques are highly desired in deep learning for alleviating overfitting and improving generalization. This work proposes a new regularization scheme, based on the understanding that the flat local minima of the empirical risk cause the model to generalize better. This scheme is referred to as adversarial model perturbation (AMP), where instead of directly minimizing the empirical risk, an alternative "AMP loss" is minimized via SGD. Specifically, the AMP loss is obtained from the empirical risk by applying the "worst" norm-bounded perturbation on each point in the parameter space. Comparing with most existing regularization schemes, AMP has strong theoretical justifications, in that minimizing the AMP loss can be shown theoretically to favour flat local minima of the empirical risk. Extensive experiments on various modern deep architectures establish AMP as a new state of the art among regularization schemes. Our code is available at https://github.com/hiyouga/AMP-Regularizer.

Aspect-level sentiment classification (ALSC) aims at predicting the sentiment polarity of a specific aspect term occurring in a sentence. This task requires learning a representation by aggregating the relevant contextual features concerning the aspect term. Existing methods cannot sufficiently leverage the syntactic structure of the sentence, and hence are difficult to distinguish different sentiments for multiple aspects in a sentence. We perceive the limitations of the previous methods and propose a hypothesis about finding crucial contextual information with the help of syntactic structure. For this purpose, we present a neural network model named RepWalk which performs a replicated random walk on a syntax graph, to effectively focus on the informative contextual words. Empirical studies show that our model outperforms recent models on most of the benchmark datasets for the ALSC task. The results suggest that our method for incorporating syntactic structure enriches the representation for the classification.

Supervised learning relies heavily on readily available labelled data to infer an effective classification function. However, proposed methods under the supervised learning paradigm are faced with the scarcity of labelled data within domains, and are not generalized enough to adapt to other tasks. Transfer learning has proved to be a worthy choice to address these issues, by allowing knowledge to be shared across domains and tasks. In this paper, we propose two transfer learning methods Anchored Model Transfer (AMT) and Soft Instance Transfer (SIT), which are both based on multi-task learning, and account for model transfer and instance transfer, and can be combined into a common framework. We demonstrate the effectiveness of AMT and SIT for aspect-level sentiment classification showing the competitive performance against baseline models on benchmark datasets. Interestingly, we show that the integration of both methods AMT+SIT achieves state-of-the-art performance on the same task.