XPXinlong Pan

Xinlong Pan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xinlong Pan has a total of 40 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Explainable Recognition of Complex Flight Maneuvers via Retrieval-Augmented Large Language Models

    by Liqiang Ren, Haipeng Wang, Xinlong Pan, Tiantian Tang, Hongdong Wan

    Complex flight maneuver recognition (FMR) underpins intelligent flight training, including training assessment, pilot skill profiling, and flight safety monitoring. Existing FMR methods typically require large labeled datasets, generalize poorly across aircraft, and provide limited decision transparency. We propose TableManeuver, an explainable LLM-based FMR method that reformulates multivariate flight parameter time series as table-understanding inputs. The method updates no base LLM parameters and uses a small labeled training set only as a retrieval library; it is therefore not a zero-shot setting. TableManeuver first converts flight parameter sequences into tabular text that preserves temporal indices and channel semantics, reducing the mismatch between numerical time series and the textual semantic space of LLMs. It then combines domain knowledge, neighborhood sample references, and task decomposition prompts in a retrieval-augmented reasoning architecture that guides explicit step-by-step inference. We evaluate the method on a flight dataset collected from human pilots on a high-fidelity flight simulation platform. Without base LLM parameter updates, TableManeuver achieves 96.2% precision, 96.8% recall, and a 96.5% F1 score, exceeding the strongest supervised baseline by 3.5 percentage points in F1. In cross-aircraft evaluation, the F1 score decreases by only 1.4 percentage points, which is substantially smaller than the degradation observed for deep learning baselines. Retrieval-only baselines that transfer neighbor labels without LLM inference perform markedly worse, indicating that the performance gains are not explained by neighbor label transfer alone. TableManeuver combines recognition accuracy, cross-aircraft robustness, and readable step-by-step reasoning evidence, offering a practical route for applying LLMs to aviation time series analysis.

  • Open Access

    Error-Propagation-Free Learned Video Compression With Dual-Domain Progressive Temporal Alignment

    by Han Li, Shaohui Li, Wenrui Dai, Chenglin Li, Xinlong Pan, Haipeng Wang, Junni Zou, Hongkai Xiong

    Existing frameworks for learned video compression suffer from a dilemma between inaccurate temporal alignment and error propagation for motion estimation and compensation (ME/MC). The separate-transform framework employs distinct transforms for intra-frame and inter-frame compression to yield impressive rate-distortion (R-D) performance but causes evident error propagation, while the unified-transform framework eliminates error propagation via shared transforms but is inferior in ME/MC in shared latent domains. To address this limitation, in this paper, we propose a novel unifiedtransform framework with dual-domain progressive temporal alignment and quality-conditioned mixture-of-expert (QCMoE) to enable quality-consistent and error-propagation-free streaming for learned video compression. Specifically, we propose dualdomain progressive temporal alignment for ME/MC that leverages coarse pixel-domain alignment and refined latent-domain alignment to significantly enhance temporal context modeling in a coarse-to-fine fashion. The coarse pixel-domain alignment efficiently handles simple motion patterns with optical flow estimated from a single reference frame, while the refined latent-domain alignment develops a Flow-Guided Deformable Transformer (FGDT) over latents from multiple reference frames to achieve long-term motion refinement (LTMR) for complex motion patterns. Furthermore, we design a QCMoE module for continuous bit-rate adaptation that dynamically assigns different experts to adjust quantization steps per pixel based on target quality and content rather than relies on a single quantization step. QCMoE allows continuous and consistent rate control with appealing R-D performance. Experimental results show that the proposed method achieves competitive R-D performance compared with the state-of-the-arts, while successfully eliminating error propagation.

  • Chinese Medical Named Entity Recognition Enhanced by Large Language Model

    by Jizhao Zhu, Xiaolin Lv, Xinlong Pan, Zhenqiu Zhu, Chunlong Fan

  • An Entity Alignment Algorithm Based on TransR and Multi-Semantic Feature Fusion

    by Qingyun Yang, Jizhao Zhu, Zhenqiu Zhu, Xinlong Pan, Chunlong Fan

  • Open Access

    Using Psycholinguistic Clues to Index Deep Semantic Evidences: Personality Detection in Social Media Texts

    by Qirui Tang, Wenkang Jiang, Xinlong Pan, Lei Lin, Jizhao Zhu, Yihua Du, Donghong Sun

    Detecting personalities in social media content is an important application of personality psychology. Most early studies apply a coherent piece of writing to personality detection, but today, the challenge is to identify dominant personality traits from a series of short, noisy social media posts. To this end, recent studies have attempted to individually encode the deep semantics of posts, often using attention-based methods, and then relate them, or directly assemble them into graph structures. However, due to the inherently disjointed and noisy nature of social media content, constructing meaningful connections remains challenging. While such methods rely on well-defined relationships between posts, effectively capturing these connections in fragmented and sparse content is non-trivial, particularly under limited supervision or noisy input. To tackle this, we draw inspiration from the scanning reading technique—commonly recommended for efficiently processing large volumes of information—and propose an index attention mechanism as a solution. This mechanism leverages prior psycholinguistic knowledge as an “index” to guide attention, thereby enabling more effective information fusion across scattered semantic signals. Building on this idea, we introduce the Index Attention Network (IAN)—a novel framework designed to infer personality labels by performing targeted information fusion over deep semantic representations of individual posts. Through a series of experiments, IAN achieved state-of-the-art performance on the Kaggle dataset and performance comparable to graph convolutional networks (GCN) on the Pandora dataset. Notably, IAN delivered an average improvement of 13% in terms of macro-F1 scores with the Kaggle dataset. The code for IAN is available at GitHub: https://github.com/Once2gain/IAN.


  • Open Access

    Quantitative Evaluation Method for Anomaly Levels of Complex Flight Maneuver Based on Multi-sensor Data

    by Liqiang Ren, Haipeng Wang, Xinlong Pan, Shuyi Jia, Bing Wan

    The methods that identify complex flight maneuvers from multi-sensor flight parameter data fusion and conduct automated quantitative evaluations of anomaly levels could play an important role in enhancing flight safety and pilot training. However, existing methods focus on anomaly detection at individual flight parameter data points, making it challenging to accurately quantify the overall abnormality of a flight maneuver. To address this issue, this paper proposes a novel method for the quantitative evaluation of anomaly levels in complex flight maneuvers using multi-sensor data. The proposed method comprises two stages: complex flight maneuver recognition and anomaly level quantification. In the complex flight maneuver recognition stage, a one-dimensional dual attention mechanism (1D-DAM) is introduced to capture discriminative fusion features in both the temporal and variable dimensions. Based on this mechanism, we develop a one-dimensional dual attention mechanism ResNet (1D-DAMResNet) model to achieve the recognition of complex flight maneuvers. Subsequently, in the anomaly level quantification stage, we employ a clustering technique to establish a standard maneuver benchmark library, which serves as a reference for the flight maneuver evaluations of different categories. According to the results of flight maneuver recognition, the corresponding category of standard maneuver from the library is automatically selected, and the dynamic time warping algorithm is then utilized to compute the anomaly quantification score of the test maneuver, thereby determining its anomaly level. Compared to contrastive methods, the proposed complex flight maneuver recognition model demonstrates significant advantages in both accuracy and stability, with an average precision, recall, and F1 scores of 99.75%. Additionally, the proposed anomaly level quantification method provides an automatic quantification of the overall anomaly level of maneuvers, and the results are highly interpretable. Overall, this paper introduces a novel approach for the quantitative evaluation of anomaly levels in maneuvers, which not only contributes to improving the accuracy of flight training evaluation but also significantly enhances the efficiency and quality of flight training.


  • Open Access

    OGSS: An Ontology-Guided and Scheduled-Sampling Approach for Overlapping Event Extraction

    by Jizhao Zhu, Hualong Wen, Xinlong Pan, Xiang Li

    Event extraction is a complex and challenging task in the field of information extraction. It aims to identify event types, triggers, and argument information from the text. In recent years, overlapping event extraction has attracted the attention of researchers because of its higher challenge and practicability, and some work has carried out in-depth research on overlapping event extraction and achieved remarkable results. But these works (1) ignore the role of ontology knowledge in event extraction; (2) use the same semantic encoding for multi-stage models, lacking consideration for the independent characteristics of extraction tasks such as event types, triggers, and arguments; and (3) face issues in the training process of multi-stage models, such as error cascading and slow convergence. To address the above issues, we propose an ontology-guided and scheduled-sampling approach for overlapping event extraction, termed as OGSS. First, we design a symmetric matrix for event ontology knowledge representation and integrate it into the semantic encoding process, infusing ontology knowledge into event extraction. Second, for extraction targets such as event types, triggers, and arguments, we process the semantic encoding according to the characteristics of each extraction target, obtaining semantic representations tailored for each subtask. Finally, we view multi-stage predictions as sequential outputs of a joint model, using a scheduled sampling strategy between subtasks to effectively mitigate the cascading propagation of errors during training and accelerate model convergence. We conduct extensive experiments on the FewFc event extraction benchmark dataset. The results show that OGSS achieves significant improvements in overlapping event extraction tasks compared to previous methods.


  • Open Access

    ES4ED: An Event Detection Model Based on Event Sentence Pre-Determination

    by Jizhao Zhu, Haonan Zhao, Wenyu Duan, Xinlong Pan, Chunlong Fan

    As an important task in the field of information extraction, event detection is widely used in event graph construction and network public opinion monitoring. Although the existing methods (such as BGCN, MGRN-EE, etc.) have obtained well performance on event detection by utilizing various features from text, they neglect that the events in data follows a long-tailed distribution, which leads to a serious bias in the trained event detection model. To address this issue, we propose an event detection model based on event sentence pre-determination, termed as ES4ED, in a simple but effective way. The model first employs classification method to identify the sentences that contain events semantically (called event sentences), and then conducts event detection on these event sentences to solve the long-tailed distribution of events. ES4ED consists of three components: the semantic encoder, the event sentence decider and the event detector. First, the semantic encoder encodes the words semantically. Then, the event sentence decider identifies event sentences by classification. Finally, the event sentences are input to the event detector to complete the event triggers identification and classification. Experimental results on the public dataset ACE2005 show that the F1 score of the proposed model achieves 79.2% and 76.5% on trigger identification and trigger classification, respectively, which are significantly improved compared with the existing typical works.


  • Research on Improved Image Dehazing Algorithm Based on Dark Channel Prior

    by Xinlong Pan, Haipeng Wang, Yu Liu, Ziwei Zhao, Fan Huang

  • Design and Simulation of Heading Controller for Unmanned Boat Based on Fuzzy Neural PID

    by Yunpeng Su, Yong Chen, Xinlong Pan, Haipeng Wang, Ziwei Zhao, Hao Liu