Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


He Ma has a total of 10 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Building and Validating a Coal Mine Safety Question-Answering System with a Large Language Model Through a Two-Stage Fine-Tuning Method

    by Zongyu Li, Xingli Liu, Shiqun Liu, He Ma, Gang Wu

    Artificial intelligence technology holds significant importance for building intelligent question-answering systems in the field of coal mine safety and enhancing safety management levels. Currently, there is a lack of specialized large language models and high-quality question-answering datasets in this field. To address this, this study proposes a two-stage fine-tuning method based on Low-Rank Adaptation (LoRA) and Group Sequence Policy Optimization (GSPO) for training a question-answering model tailored to the coal mine safety domain. The research begins by constructing a dedicated question-answering dataset based on domain-specific regulatory documents. Subsequently, using Qwen2.5-7B Instruct as the base model, the study fine-tunes the model through supervised learning with LoRA technology, followed by further optimization of the model’s performance using the GSPO reinforcement learning algorithm. Experiments show that the model trained with this method exhibits significant improvements in coal mine safety-related tasks, achieving superior results on multiple automated evaluation metrics compared to contrast models of similar scale. This study validates the effectiveness of the two-stage fine-tuning method in adapting large language models (LLMs) to specific domains, providing a new technical approach for the intelligentization of coal mine safety. It should be noted that due to the lack of external data, this study relies on a self-constructed dataset and has not yet undergone external independent validation, which constitutes the main limitation of the current work.


  • Open Access

    Research on Large Language Model-Based Automatic Knowledge Extraction for Coal Mine Equipment Safety

    by Ziheng Zhang, Rijia Ding, Yinhang Liu, He Ma

    Structured knowledge representation is of great significance for constructing a knowledge graph of coal mine equipment safety. However, traditional methods encounter substantial difficulties when handling the complex semantics and domain-specific terms in technical texts. To tackle this challenge, we propose a knowledge extraction framework that integrates large language models (LLMs) with prompt engineering to achieve the efficient joint extraction of information. This framework strengthens the traditional triple structure by introducing symmetric entity-type information encompassing the head entity type and the tail entity type. Furthermore, it enables simultaneous entity recognition and relation extraction within a unified model. Experimental results demonstrate that the proposed knowledge extraction framework significantly outperforms the traditional step-by-step approach of first extracting entities and then relations. To meet the requirements of actual industrial production, we verified the impacts of different prompt strategies, as well as small lightweight models and large complex models, on the extraction task. Through multiple sets of comparative experiments, we found that the Chain-of-Thought (CoT) prompting strategy can effectively improve performance across different models, and the choice of model architecture has a significant impact on task performance. Our research provides an accurate and scalable solution for knowledge graph construction in the coal mine equipment safety domain, and its symmetry-aware design exhibits great potential for cross-domain knowledge transfer.


  • Open Access

    A deep clustering framework integrating pairwise constraints and a VMF mixture model

    by He Ma, Weipeng Wu

    We presented a novel deep generative clustering model called Variational Deep Embedding based on Pairwise constraints and the Von Mises-Fisher mixture model (VDEPV). VDEPV consists of fully connected neural networks capable of learning latent representations from raw data and accurately predicting cluster assignments. Under the assumption of a genuinely non-informative prior, VDEPV adopted a von Mises-Fisher mixture model to depict the hyperspherical interpretation of the data. We defined and established pairwise constraints by employing a random sample mining strategy and applying data augmentation techniques. These constraints enhanced the compactness of intra-cluster samples in the spherical embedding space while improving inter-cluster samples' separability. By minimizing Kullback-Leibler divergence, we formulated a clustering loss function based on pairwise constraints, which regularized the joint probability distribution of latent variables and cluster labels. Comparative experiments with other deep clustering methods demonstrated the excellent performance of VDEPV.


  • Open Access

    Achieving deep clustering through the use of variational autoencoders and similarity-based loss

    by He Ma

    Clustering is an important and challenging research topic in many fields. Although various clustering algorithms have been developed in the past, traditional shallow clustering algorithms cannot mine the underlying structural information of the data. Recent advances have shown that deep clustering can achieve excellent performance on clustering tasks. In this work, a novel variational autoencoder-based deep clustering algorithm is proposed. It treats the Gaussian mixture model as the prior latent space and uses an additional classifier to distinguish different clusters in the latent space accurately. A similarity-based loss function is proposed consisting specifically of the cross-entropy of the predicted transition probabilities of clusters and the Wasserstein distance of the predicted posterior distributions. The new loss encourages the model to learn meaningful cluster-oriented representations to facilitate clustering tasks. The experimental results show that our method consistently achieves competitive results on various data sets.