While Large Language Models (LLMs) have demonstrated exceptional multitasking abilities, fine-tuning these models on downstream, domain-specific datasets is often necessary to yield superior performance on test sets compared to their counterparts without fine-tuning. However, the comprehensive effects of fine-tuning on the LLMs' generalization ability are not fully understood. This paper delves into the differences between original, unmodified LLMs and their fine-tuned variants. Our primary investigation centers on whether fine-tuning affects the generalization ability intrinsic to LLMs. To elaborate on this, we conduct extensive experiments across five distinct language tasks on various datasets. Our main findings reveal that models fine-tuned on generation and classification tasks exhibit dissimilar behaviors in generalizing to different domains and tasks. Intriguingly, we observe that integrating the in-context learning strategy during fine-tuning on generation tasks can enhance the model's generalization ability. Through this systematic investigation, we aim to contribute valuable insights into the evolving landscape of fine-tuning practices for LLMs.
Decoding methods play an indispensable role in converting language models from next-token predictors into practical task solvers. Prior research on decoding methods, primarily focusing on task-specific models, may not extend to the current era of general-purpose large language models (LLMs). Moreover, the recent influx of decoding strategies has further complicated this landscape. This paper provides a comprehensive and multifaceted analysis of various decoding methods within the context of LLMs, evaluating their performance, robustness to hyperparameter changes, and decoding speeds across a wide range of tasks, models, and deployment environments. Our findings reveal that decoding method performance is notably task-dependent and influenced by factors such as alignment, model size, and quantization. Intriguingly, sensitivity analysis exposes that certain methods achieve superior performance at the cost of extensive hyperparameter tuning, highlighting the trade-off between attaining optimal results and the practicality of implementation in varying contexts.
Hypergraphs provide an effective modeling approach for modeling high-order relationships in many real-world datasets. To capture such complex relationships, several hypergraph neural networks have been proposed for learning hypergraph structure, which propagate information from nodes to hyperedges and then from hyperedges back to nodes. However, most existing methods focus on information propagation between hyperedges and nodes, neglecting the interactions among hyperedges themselves. In this paper, we propose HeIHNN, a hyperedge interaction-aware hypergraph neural network, which captures the interactions among hyperedges during the convolution process and introduce a novel mechanism to enhance information flow between hyperedges and nodes. Specifically, HeIHNN integrates the interactions between hyperedges into the hypergraph convolution by constructing a three-stage information propagation process. After propagating information from nodes to hyperedges, we introduce a hyperedge-level convolution to update the hyperedge embeddings. Finally, the embeddings that capture rich information from the interaction among hyperedges will be utilized to update the node embeddings. Additionally, we introduce a hyperedge outlier removal mechanism in the information propagation stages between nodes and hyperedges, which dynamically adjusts the hypergraph structure using the learned embeddings, effectively removing outliers. Extensive experiments conducted on real-world datasets show the competitive performance of HeIHNN compared with state-of-the-art methods.
Recent studies have shown that attackers can catastrophically reduce the performance of GNNs by maliciously modifying the graph structure or node features on the graph. Adversarial training, which has been shown to be one of the most effective defense mechanisms against adversarial attacks in computer vision, holds great promise for enhancing the robustness of GNNs. There is limited research on defending against attacks by performing adversarial training on graphs, and it is crucial to delve deeper into this approach to optimize its effectiveness. Therefore, based on robust adversarial training on graphs, we propose a hierarchical constraint refinement framework (HC-Ref) that enhances the anti-perturbation capabilities of GNNs and downstream classifiers separately, ultimately leading to improved robustness. We propose corresponding adversarial regularization terms that are conducive to adaptively narrowing the domain gap between the normal part and the perturbation part according to the characteristics of different layers, promoting the smoothness of the predicted distribution of both parts. Moreover, existing research on graph robust adversarial training primarily concentrates on training from the standpoint of node feature perturbations and seldom takes into account alterations in the graph structure. This limitation makes it challenging to prevent attacks based on topological changes in the graph. This paper generates adversarial examples by utilizing graph structure perturbations, offering an effective approach to defend against attack methods that are based on topological changes. Extensive experiments on two real-world graph benchmarks show that HC-Ref successfully resists various attacks and has better node classification performance compared to several baseline methods.
Large language models (LLMs) have shown surprisingly good performance in multilingual neural machine translation (MNMT) even when trained without parallel data. Yet, despite the fact that the amount of training data is gigantic, they still struggle with translating rare words, particularly for low-resource languages. Even worse, it is usually unrealistic to retrieve relevant demonstrations for in-context learning with low-resource languages on LLMs, which restricts the practical use of LLMs for translation -- how should we mitigate this problem? To this end, we present a novel method, CoD, which augments LLMs with prior knowledge with the chains of multilingual dictionaries for a subset of input words to elicit translation abilities for LLMs. Extensive experiments indicate that augmenting ChatGPT with CoD elicits large gains by up to 13x chrF++ points for MNMT (3.08 to 42.63 for English to Serbian written in Cyrillic script) on FLORES-200 full devtest set. We further demonstrate the importance of chaining the multilingual dictionaries, as well as the superiority of CoD to few-shot demonstration for low-resource languages.
We introduce a frustratingly simple, super efficient and surprisingly effective decoding method, which we call Frustratingly Simple Decoding (FSD), for neural text generation. The idea behind FSD is straightforward: we build an anti-LM based on previously generated text and use this anti-LM to penalize future generation of what has been generated. The anti-LM can be implemented as simple as an n-gram language model or a vectorized variant. In this way, FSD introduces no extra model parameters and negligible computational overhead (FSD can be as fast as greedy search). Despite the simplicity, FSD is surprisingly effective; Experiments show that FSD can outperform the canonical methods to date (i.e., nucleus sampling) as well as several strong baselines that were proposed recently.
Fine-tuning pre-trained models has been ubiquitously proven to be effective in a wide range of NLP tasks. However, fine-tuning the whole model is parameter inefficient as it always yields an entirely new model for each task. Currently, many research works propose to only fine-tune a small portion of the parameters while keeping most of the parameters shared across different tasks. These methods achieve surprisingly good performance and are shown to be more stable than their corresponding fully fine-tuned counterparts. However, such kind of methods is still not well understood. Some natural questions arise: How does the parameter sparsity lead to promising performance? Why is the model more stable than the fully fine-tuned models? How to choose the tunable parameters? In this paper, we first categorize the existing methods into random approaches, rule-based approaches, and projection-based approaches based on how they choose which parameters to tune. Then, we show that all of the methods are actually sparse fine-tuned models and conduct a novel theoretical analysis of them. We indicate that the sparsity is actually imposing a regularization on the original model by controlling the upper bound of the stability. Such stability leads to better generalization capability which has been empirically observed in a lot of recent research works. Despite the effectiveness of sparsity grounded by our theory, it still remains an open problem of how to choose the tunable parameters. To better choose the tunable parameters, we propose a novel Second-order Approximation Method (SAM) which approximates the original problem with an analytically solvable optimization function. The tunable parameters are determined by directly optimizing the approximation function. The experimental results show that our proposed SAM model outperforms many strong baseline models and it also verifies our theoretical analysis.
The rapid advances in automation technologies, such as artificial intelligence (AI) and robotics, pose an increasing risk of automation for occupations, with a likely significant impact on the labour market. Recent social-economic studies suggest that nearly 50\% of occupations are at high risk of being automated in the next decade. However, the lack of granular data and empirically informed models have limited the accuracy of these studies and made it challenging to predict which jobs will be automated. In this paper, we study the automation risk of occupations by performing a classification task between automated and non-automated occupations. The available information is 910 occupations' task statements, skills and interactions categorised by Standard Occupational Classification (SOC). To fully utilize this information, we propose a graph-based semi-supervised classification method named \textbf{A}utomated \textbf{O}ccupation \textbf{C}lassification based on \textbf{G}raph \textbf{C}onvolutional \textbf{N}etworks (\textbf{AOC-GCN}) to identify the automated risk for occupations. This model integrates a heterogeneous graph to capture occupations' local and global contexts. The results show that our proposed method outperforms the baseline models by considering the information of both internal features of occupations and their external interactions. This study could help policymakers identify potential automated occupations and support individuals' decision-making before entering the job market.
Graph learning models are critical tools for researchers to explore graph-structured data. To train a capable graph learning model, a conventional method uses sufficient training data to train a graph model on a single device. However, it is prohibitive to do so in real-world scenarios due to privacy concerns. Federated learning provides a feasible solution to address such limitations via introducing various privacy-preserving mechanisms, such as differential privacy on graph edges. Nevertheless, differential privacy in federated graph learning secures the classified information maintained in graphs. It degrades the performances of the graph learning models. In this paper, we investigate how to implement differential privacy on graph edges and observe the performances decreasing in the experiments. We also note that the differential privacy on graph edges introduces noises to perturb graph proximity, which is one of the graph augmentations in graph contrastive learning. Inspired by that, we propose to leverage the advantages of graph contrastive learning to alleviate the performance dropping caused by differential privacy. Extensive experiments are conducted with several representative graph models and widely-used datasets, showing that contrastive learning indeed alleviates the models' performance dropping caused by differential privacy.