• Plang: A language blending natural prompt with control flow for precise LLM guidance. • Meta-prompt programming: Enables LLM to self-modify prompt programs during execution. • 40.75%-89.55% conciseness gain: Outperforms methods like LangChain in prompt coding. • Open-source solution: Enables multi-agent collaboration & tool use with minimal code. The advent of instruction-following large language models (LLMs), exemplified by ChatGPT, has significantly enhanced the performance of generative autoregressive natural language models on general tasks, marking a crucial milestone toward artificial general intelligence. While research has shown that LLM performance critically depends on prompt effectiveness, existing prompt construction approaches - including prompt string templates, prompt programming frameworks, and prompt programming syntactic sugar - suffer from limitations in imprecise generation control and deviation from natural language syntax, thereby impeding prompt engineering advancement. To address these challenges, we introduce PromptLanguage (Plang), a string-first programming language designed specifically for LLM prompt engineering. Our key innovation lies in utilizing font styles as syntax keywords to seamlessly integrate natural language prompt text with control flow code, enabling precise intervention in the generation process while maintaining the natural language affinity of prompt programs. Notably, Plang pioneers the concept of meta-prompt programming. Extensive experimental results across prompt engineering cases and quantitative analyses demonstrate that Plang-written prompt programs offer superior read/writability, intervention precision and up to 89.55% efficiency improvements. The Plang implementation is available as open-source software at https://github.com/HJZ-XDU/plang .

Excessively large classes that encapsulate multiple responsibilities are challenging to comprehend and maintain. Addressing this issue, several Extract Class refactoring tools have been proposed, employing a two-phase process: identifying suitable fields or methods for extraction, and implementing the mechanics of refactoring. These tools traditionally generate an intra-class dependency graph to analyze the class structure, applying hard-coded rules based on this graph to unearth refactoring opportunities. Yet, the graph-based approach predominantly illuminates direct, “one-to-one” relationship between pairwise entities. Such a perspective is restrictive as it overlooks the complex, “one-to-many” dependencies among multiple entities that are prevalent in real-world classes. This narrow focus can lead to refactoring suggestions that may diverge from developers’ actual needs, given their multifaceted nature. To bridge this gap, our paper leverages the concept of intra-class dependency hypergraph to model one-to-many dependency relationship and proposes a hypergraph learning-based approach to suggest Extract Class refactoring opportunities named HECS. For each target class, we first construct its intra-class dependency hypergraph and assign attributes to nodes with a pre-trained code model. All the attributed hypergraphs are fed into an enhanced hypergraph neural network for training. Utilizing this trained neural network alongside a large language model (LLM), we construct a refactoring suggestion system. We trained HECS on a large-scale dataset and evaluated it on two real-world datasets. The results show that demonstrates an increase of 38.5% in precision, 9.7% in recall, and 44.4% in f1-measure compared to 3 state-of-the-art refactoring tools including JDeodorant, SSECS, and LLMRefactor, which is more useful for 64% of participants. The results also unveil practical suggestions and new insights that benefit existing extract-related refactoring techniques.

Methods implemented in incorrect classes will cause excessive reliance on other classes than their own, known as a typical code smell symptom: feature envy, which makes it difficult to maintain increased coupling between classes. Addressing this issue, several Move Method refactoring tools have been proposed, employing a two-phase process: identifying misplaced methods to move and appropriate classes to receive, and implementing the mechanics of refactoring. These tools traditionally use hard-coded metrics to measure correlations between movable methods and target classes and apply heuristic thresholds or trained classifiers to unearth refactoring opportunities. Yet, these approaches predominantly illuminate pairwise correlations between methods and classes while overlooking the complex and complicated dependencies binding multiple code entities within these methods/classes that are prevalent in real-world cases. This narrow focus can lead to refactoring suggestions that may diverge from developers' actual needs. To bridge this gap, our paper leverages the concept of inter-class code entity dependency hypergraph to model complicated dependency relationships involving multiple code entities within various methods/classes and proposes a hypergraph learning-based approach to suggest Move Method refactoring opportunities named HMove. We first construct inter-class code entity dependency hypergraphs from training samples and assign attributes to entities with a pre-trained code model. All the attributed hypergraphs are fed into a hybrid hypergraph neural network for training. Utilizing this trained neural network alongside a large language model, we construct a refactoring suggestion system. We trained HMove on a large-scale dataset and evaluated it on two real-world datasets. The results show that demonstrates an increase of 27.8% in precision, 2.5% in recall, and 18.5% in f1-measure compared to 9 state-of-the-art refactoring tools, which is more useful for 68% of participants. The results also unveil practical suggestions and new insights that benefit existing feature envy-related refactoring techniques.

Deep-learning based methods that aim to extract effective high-level features have steadily improved the performance on the speech emotion recognition. However, low-level features that contain important emotion-related information have not gained much attention. In this paper, we propose a novel low-level feature extraction method based on the Time-Frequency Attention (TFA) module and Time-Frequency Weighting (TFW) module. First, the TFA module is designed to learn notable regions in the detail-rich low-level feature maps produced by the scale-specific convolutional layers. Then, the TFW module is proposed to extract discriminative features from the time and frequency dimensions respectively. Finally, the speech emotion recognition task is completed by the subsequent multi-branch network. Experimental results on the IEMOCAP and RAVDESS datasets demonstrate the importance of low-level features, and show that the proposed method outperforms other state-of-the-art approaches.

In intelligent human-computer interaction systems, speech emotion recognition (SER) is a fundamental task for understanding user intention. One vital challenge for emotion inferring is how to extract discriminative and robust features. In this paper, we propose a novel network based on the Time-Frequency Weighting (TFW) module and the ConvlD enabled Multi-head Element-wise Self-attention (ID-MESA) block to extract discriminative features from three dimensions of time, frequency and channel for improving the performance in SER. The TFW module is designed to capture emotion information along the time and frequency dimensions in the shallow neural network. As the high complexity of the emotion feature, the 1D-MESA block can assist the network to locate the discriminative emotion features in the channel dimension. The proposed architecture outperforms the state-of-the-art methods in the IEMOCAP database, with the absolute increase of 3.98% and 1.58% on unweighted accuracy among four emotion classes and weighted accuracy, respectively.

Abstract In recent years, Deep Neural Networks (DNNs) have achieved excellent performance on many tasks, but it is very difficult to train good models from imbalanced datasets. Creating balanced batches either by majority data down-sampling or by minority data up-sampling can solve the problem in certain cases. However, it may lead to learning process instability and overfitting. In this paper, we propose the Batch Balance Wrapper (BBW), a novel framework which can adapt a general DNN to be well trained from extremely imbalanced datasets with few minority samples. In BBW, two extra network layers are added to the start of a DNN. The layers prevent overfitting of minority samples and improve the expressiveness of the sample distribution of minority samples. Furthermore, Batch Balance (BB), a class-based sampling algorithm, is proposed to make sure the samples in each batch are always balanced during the learning process. We test BBW on three well-known extremely imbalanced datasets with few minority samples. The maximum imbalance ratio reaches 1167:1 with only 16 positive samples. Compared with existing approaches, BBW achieves better classification performance. In addition, BBW-wrapped DNNs are 16.39 times faster, relative to unwrapped DNNs. Moreover, BBW does not require data preprocessing or additional hyper-parameter tuning, operations that may require additional processing time. The experiments prove that BBW can be applied to common applications of extremely imbalanced data with few minority samples, such as the classification of EEG signals, medical images and so on.

Convolutional Neural Networks (CNNs) have achieved remarkable performance breakthroughs in a variety of tasks. Recently, CNN-based methods that are fed with hand-extracted EEG features have steadily improved their performance on the emotion recognition task. In this paper, we propose a novel convolutional layer, called the Scaling Layer, which can adaptively extract effective data-driven spectrogram-like features from raw EEG signals. Furthermore, it exploits convolutional kernels scaled from one data-driven pattern to exposed a frequency-like dimension to address the shortcomings of prior methods requiring hand-extracted features or their approximations. ScalingNet, the proposed neural network architecture based on the Scaling Layer, has achieved state-of-the-art results across the established DEAP and AMIGOS benchmark datasets.

Air pollution is a major hazard to environment and human health. Particle matter with a diameter less than 2.5 micrometers (PM25) is a very harmful air pollutant that can penetrate deeply into lungs through human respiratory system. In this paper, we propose an efficient and reliable method to estimate PM25concentration using outdoor images. Firstly, a prior attention block based on gradient features is used to enhance the boundary area between the sky region and the object in a feature map. After that, an embedding layer is applied to encode weather information and fuse it with image features. Finally, a deep neural network model with a novel loss function, F-MSE, is constructed to combine the prediction error of each model level during the training process and to further improve the effectiveness of the presented method. The proposed method was evaluated on a PM2.5dataset with 1,514 images and the experimental results demonstrate that our method outperformed other existing methods.