Acoustic scene classification under lowcomplexity conditions presents unique challenges due to limited data availability and computational constraints. To address this, we propose a lightweight and effective ASC framework that integrates multi-scale feature fusion with attention mechanisms, specifically designed for resource-constrained environments. The model begins by extracting time-frequency representations using 3D convolutional layers, which efficiently capture spatial and temporal patterns in audio signals. To further emphasize critical information, an attention mechanism guided by timesemantic neighbors dynamically assigns three-dimensional weights to salient features, highlighting the most informative components in the time-frequency space. In addition, a pyramid pooling module with multi-scale dilated convolutions is employed to aggregate both local and global contextual cues, enhancing the model's ability to detect distinctive acoustic events at specific frequency ranges. This hierarchical design improves feature expressiveness without introducing significant computational overhead. Extensive experiments on multiple benchmark datasets demonstrate that our method not only achieves competitive or superior classification accuracy compared to state-of-the-art techniques, but also maintains low model complexity and inference cost, making it highly suitable for deployment in real-world, low-complexity scenarios.

Legal Similar Case Retrieval (LSCR) is a critical application in legal Artificial Intelligence (AI). It involves retrieving the most relevant cases from legal case databases through query cases. Legal cases are semi-structured documents characterized by long text sequences and high specialization. Existing approaches rely on pre-trained language models for retrieval. However, these methods are constrained by the length of text input, preventing them from fully comprehending cases, which results in poor retrieval performance. Knowledge Graph (KG) are a type of graph structure data with dense knowledge and clear logic, which can represent the criminal processes and relationships among characters within legal cases. Currently, the mainstream approach for handling KG remains Graph Neural Networks (GNNs). However, these methods are limited by message passing and are prone to over-smoothing problems in the process of aggregating node features. To address these issues, we propose a Legal similar case retrieval model that combines Graph representation learning with the Transformer, called LeGalFormer. Three encoding methods are introduced to incorporate the structural information of the graph into the Transformer architecture. We evaluate the model on a real legal dataset, and the experimental results show that LeGalFormer significantly enhances the model's understanding capacity and achieves state-of-the-art performance.

Legal Judgment Prediction (LJP), a paramount application of artificial intelligence in law, aims to infer outcomes from fact descriptions. These outcomes mainly involve determining applicable law articles, charges, and the term of penalty. Existing approaches typically treat this task as a text classification problem, utilizing end-to-end black box models for automatic prediction. However, these approaches overlook the fine-grained elements within cases and limit the model’s interpretability. In light of these limitations, we propose a novel approach for LJP. Our method employs Fine-Grained Element Graphs (FEGs) to represent fact descriptions and applies Graph Attention Network (GAT) to learn graph representation for subsequent prediction tasks. FEG is characterized by capturing the relevant adjudicative elements of each fact description and preserving important relations among them, such as causality and temporal sequences while filtering out irrelevant content. It enhances LJP task performance and provides interpretable references for legal professionals. Since increased or decreased penalties corresponding to different elements cannot be obtained from fact descriptions, we introduce Sentencing Guidance (SG) and combine it with FEGs as a supplementary feature. Experiments confirm the effectiveness of our method. Specifically, compared to the best baseline model, it improves macro-F1 by 3.09%, 0.57%, and 1.78% for the subtasks of law article prediction, charge prediction, and term of penalty prediction.

Extracting entity relations is vital in legal artificial intelligence. It automates the mining of triple data from vast legal texts. Current methods face challenges in inaccurately identifying legal named entity boundaries and extracting over-lapping relation triples from legal texts. We present KGNet, a model developed to address these issues effectively. Our approach introduces a Word Information Generator Based on BMES tagging combined with the Fusionformer module. This innovation enhances the incorporation of legal domain knowl-edge into text representations, improving the accuracy of entity recognition. Additionally, we utilize the GlobalPointer decoder, which redefines and decomposes relation triples, thus resolving the issue of overlapping entities. Performance evaluations on a specially constructed judicial document dataset show that KGNet achieves an F1 score of 66.7%, representing an average improvement of 15.3% over baseline models. These results confirm the effectiveness of KGNet in enhancing legal document processing.

The task of charge prediction aims to determine the ultimate charge of defendant by analyzing factual description in legal cases. Existing approaches mainly use the factual descriptions provided in the dataset to guide this task. However, Criminal Charge Definitions (CCDs) in laws and regulations is a rich additional information that can effectively distinguish between confusing charges, which is not considered by existing models. In this work, we present a model termed Charge Prediction Sequence Graph with External Knowledge (EK-CPSG), which integrates information about Criminal Charge Definitions (CCDs). This method is based on the charge prediction model CP-KG, which overcomes the problem of inaccurate charge prediction due to unbalanced data and incomplete extraction of key elements. Firstly, we use attention mechanism to effectively combine CCDs encoded by TextCNN with case facts encoded by Bi-GRU. Secondly, we focus on the most relevant elements of case facts and CCDs. Thirdly, to evaluate EK-CPSG, we constructed CCs-8 from CAIL2018 dataset, which includes 8 confusing charges, and conducted a large number of experiments on it. The experimental results indicate that EK-CPSG achieves scores of 90.96% in Macro-F1. Compared to the baseline model TFIDF-SVM, EK-CPSG shows significant improvements with 14.55%. When compared to CP-KG, there is an approximately 0.73% enhancement. Finally, through ablation experiments, we confirmed the rationality and reliability of the introduction of CCDs.

In this paper, we proposed a multi-layer pre-training approach to overcoming the limitations of extractive machine reading comprehension (EMRC) in capturing global semantic information in depth and facilitating deep interaction between text and questions. The proposed framework adopted the Masked-Language Model correction BERT (MacBERT) model and multi-layer perception (MLP) to predict the probability of each location as an answer. To address span-extractive, unanswerable and YES/NO questions, we employed Bi-directional long short-term memory (BiLSTM) and self-attention to create different targeted layers of the neural network model. The novel model exhibits robust generalization and interactive extraction capability. With the Chinese Judicial Reading Comprehension (CJRC) dataset, the experimental results show that the proposed algorithm brings significant performance improvement by 3.5% in civil cases and 4.9% in criminal cases in terms of F1 score.

Multi-channel signal processing techniques have played an important role in the far-field automatic speech recognition (ASR) as the separate front-end enhancement part. However, they often meet the mismatch problem. In this paper, we proposed a novel architecture of acoustic model, in which the multi-channel speech without preprocessing was utilized directly. Besides the strategy of knowledge distillation and the generalized cross correlation (GCC) adaptation were employed. We use knowledge distillation to transfer knowledge from a well-trained close-talking model to distant-talking scenarios in every frame of the multichannel distant speech. Moreover, the GCC between microphones, which contains the spatial information, is supplied as an auxiliary input to the neural network. We observe good compensation of those two techniques. Evaluated with the AMI and ICSI meeting corpora, the proposed methods achieve relative WER improvement of 7.7% and 7.5% over the model trained directly on the concatenated multi-channel speech.

This paper presents a novel far-field voice trigger algorithm utilizing DNN with the objective function of state-level minimum Bayes risk for training, customizing the decoding network to absorb the ambient noise and background speech. We adopt a two-stage classification strategy to integrate the phonetic knowledge and model-based classification into detecting wake-up words. Experimental results of the online test show that it can provide a higher than 90% accuracy and meanwhile false alarms are less than once per nine hours in the noisy home environments where the sound pressure level is about 80dB.

We propose a novel data utilization strategy, called multichannel-condition learning, leveraging upon complementary information captured in microphone array speech to jointly train dereverberation and acoustic deep neural network (DNN) models for robust distant speech recognition. Experimental results, with a single automatic speech recognition (ASR) system, on the REVERB2014 simulated evaluation data show that, on 1-channel testing, the baseline joint training scheme attains a word error rate (WER) of 7.47%, reduced from 8.72% for separate training. The proposed multi-channel-condition learning scheme has been experimented on different channel data combinations and usage showing many interesting implications. Finally, training on all 8-channel data and with DNN-based language model rescoring, a state-of-the-art WER of 4.05% is achieved. We anticipate an even lower WER when combining more top ASR systems.