Xiaoke Qi

Existing AI-based meeting summarization tools have enabled rapid generation of meeting notes, yet their reliability and user controllability remain limited. This paper explores human-AI collaboration for mobile meeting summarization and presents MeetSumAid, a multifunctional system that integrates summarization algorithms with an interactive user interface. The system is designed to support users in understanding, validating, and refining AI-generated summaries through natural interactions and flexible control mechanisms. By enabling real-time inspection, editing, and feedback, MeetSumAid facilitates reliable collaboration between humans and AI in dynamic meeting scenarios. A user study with 20 participants shows that MeetSumAid significantly improves summary quality, generation efficiency, and user-perceived reliability compared with baseline AI summarizers, while reducing cognitive load. Further analysis reveals how different interface components enhance users' engagement and confidence during collaboration. This work provides a practical step toward reliable and user-centered human-AI collaboration in mobile meeting summarization and offers actionable design implications for future intelligent collaborative systems.

Tibetan text recognition plays a key role in preserving the Tibetan language, religion, and traditions. While text recognition has made progress for high-resource languages, handwritten Tibetan character recognition remains difficult due to limited data and the lack of public large language models. Most existing datasets focus on printed or historical documents, as well as online handwriting data, but there are still few large offline handwritten Tibetan datasets. To solve this problem, we construct TibHCR, a large-scale offline handwritten character recognition dataset for the Tibetan language. To increase the diversity of the linguistic and font styles, more character categories and participants from 5 provinces in China are included. To collect and label the data efficiently, we introduce a grid sheet design, reducing manual annotation to just 1% of the samples. This design then allows for automatic data processing to extract each character sample and its corresponding label. The resulting TibHCR dataset contains 141,698 samples from 235 Tibetan writers, covering 47 character classes. We evaluate TibHCR using two recognition models: a convolutional recurrent neural network (CRNN) and a cross-lingual fine-tuning method, on a Chinese pretrained model using the PP-OCRv4 architecture to adapt Tibetan data. The results show that both models can recognize handwritten Tibetan characters efficiently, with an accuracy of 99.48% for CRNN and 99.70% for the fine-tuning method. The TibHCR dataset is publicly available at https://huggingface.co/datasets/qixiaoke/TibHCR.

The purpose of this paper is to explore the application of large language models (LLMs) in legal case retrieval and to evaluate their potential for providing legal professionals with more efficient work aids. Currently, although pre-trained models have made great progress in legal case retrieval, they are often limited to specific types of law (e.g., criminal law, civil law, etc.) and lack the ability to generalize across different types of law. Moreover, most models can only deal with a single task, whereas the legal case retrieval task requires a model to have a superb comprehension of legal texts, involving multiple subtasks and requiring multitasking capabilities. Therefore, the large language model, which has super generalization and multitasking ability, can solve the above problems. In order to explore the application of large language models for legal case retrieval in the legal domain, this paper evaluates a series of emerging large language models, including multilingual models, homegrown large models, and models specifically designed for the legal domain. These models are used to retrieve legal cases and its associated subtasks. Based on the Supreme People’s Court definition, the legal case retrieval task is broken down into seven subtasks: event detection, fact generation, trigger word extraction, keyword extraction, summarization, dispute focus identification, and reasoning generation. Using a variety of evaluation metrics, the experiments demonstrated that these emerging models have significant potential in the field of legal case retrieval, even with few shot samples. The research in this paper not only introduces new ideas in the field of legal case retrieval, but also empirically verifies the potential of LLMs to improve the quality and efficiency of retrieval. It proves the value of large language models in this field and is expected to significantly enhance the efficiency of legal practitioners, as well as promote the consistency and fairness of legal judgments through the use of emerging technologies.

While bilingual dictionaries between source and target languages can signiffcantly enhance translation performance in large language models (LLMs) by providing critical lexical alignment signals. Notably, no publicly available word-level triples exist for Chinese, Mongolian, and theirlatinized counterparts. To address this accessibility challenge, we create a new dataset that contains 30, 000 ⟨Chinese, Mongolian, Mongolian(latin)⟩ word-level triple data. More speciffcally,We ffrst leverage cross-lingual word induction to extract a Chinese-Mongolian bilingual dictionary from a large-scale and readily available Chinese-Mongolian comparable corpus. Then,with the help of ffve Mongolian language experts, we conduct quality assessment, manual proofreading, and latin transcription of the extracted dictionary to generate the triple dataset. Theopen-source release of this dataset advances Chinese-Mongolian cross-lingual information processing research while establishing a replicable framework for constructing bilingual dictionariesin other low-resource language contexts.

Low-resource text plagiarism detection faces a significant challenge due to the limited availability of labeled data for training. This task requires the development of sophisticated algorithms capable of identifying similarities and differences in texts, particularly in the realm of semantic rewriting and translation-based plagiarism detection. In this paper, we present an enhanced attentive Siamese Long Short-Term Memory (LSTM) network designed for Tibetan-Chinese plagiarism detection. Our approach begins with the introduction of translation-based data augmentation, aimed at expanding the bilingual training dataset. Subsequently, we propose a pre-detection method leveraging abstract document vectors to enhance detection efficiency. Finally, we introduce an improved attentive Siamese LSTM network tailored for Tibetan-Chinese plagiarism detection. We conduct comprehensive experiments to showcase the effectiveness of our proposed plagiarism detection framework.

Supporting a massive amount of Internet of Things applications requires a large pool of spectrum. DSM is a promising ecosystem to improve the spectrum efficiency. In the era of LoRaWAN, the physical hardware constraints, along with the bandwidth-hungry applications pose new challenges. In this article, we investigate a novel deep-reinforcement-learning-based spectrum-sharing paradigm, termed Intelligent Overlapping, that explores partially overlapping channels for concurrent spectrum access in LoRaWAN. Our key insight is to leverage the coding redundancy to expand the available spectrum without complicated data processing algorithms. In particular, we learn the extra coding redundancy from the data on the non-overlapping spectrum via a deep-Q-learning network, and we apply such redundancy to recover the data on the overlapping spectrum. In the Media Access Control layer, we predict the channel condition and strategically learn and assign the appropriate overlapping portion to the concurrent access end devices. In the Physical layer, we harness interleaving to randomize the mutual interference to ensure that all the data remains decodable. Simulation results demonstrate that Intelligent Overlapping greatly improves the spectrum efficiency with a fast convergence rate compared to the conventional DSM mechanisms.

Spatial audio has attracted more and more attention in the fields of virtual reality (VR), blind navigation and so on. The individualized head-related transfer functions (HRTFs) play an important role in generating spatial audio with accurate localization perception. Existing methods only focus on one database, and do not fully utilize the information from multiple databases. In light of this, a pre-trained-based individualization model is proposed to predict HRTFs for any target user in this paper, and a real-time spatial audio rendering system built on a wearable device is implemented to produce an immersive virtual auditory display. The proposed method first builds a pre-trained model based on multiple databases using a DNN-based model combined with an autoencoder-based dimensional reduction method. This model can capture the nonlinear relationship between user-independent HRTFs and position-dependent features. Then, fine tuning is done using a transfer learning technique at a limit number of layers based on the pre-trained model. The key idea behind fine tuning is to transfer the pre-trained user-independent model to the user-dependent one based on anthropometric features. Finally, real-time issues are discussed to guarantee a fluent auditory experience during dynamic scene update, including fine-grained head-related impulse response (HRIR) acquisition, efficient spatial audio reproduction, and parallel synthesis and playback. These techniques ensure that the system is implemented with little computational cost, thus minimizing processing delay. The experimental results show that the proposed model outperforms other methods in terms of subjective and objective metrics. Additionally, our rendering system runs on HTC Vive, with almost unnoticeable delay.

Recent years have witnessed a steep grow in the multimedia-oriented Internet of Things (IoT) over vehicular networks. Huge volume of multimedia traffic generated from the in-built IoT devices should be delivered among vehicles and immediate surroundings in real time. However, as network nodes with higher mobility, vehicles often experience more unpredictable wireless channels. Such time-frequency diversity poses substantial challenges to achieve pervasive and real-time multimedia connectivity. The hurdle lies in the inability of automatically approaching the subcarrier level channel variations in the existing video codecs. With coarse-grained traffic delivery rate, video decoding fails and intermittent connection occurs. To break this stalemate, we propose a fine-grained wireless video streaming strategy, namely iCast, that intelligently achieves the most appropriate data rate and frame protection for multimedia traffic in highly mobile vehicular environments. The insight of iCast is a simple joint source-channel rateless code. It reaps the benefits of the frequency diversity to provide fine-grained data rate for the channel in conjunction with suitable protection for the source. Our experiments show that, by harnessing frequency diversity in mobile environments, iCast outperforms the existing competitive wireless video delivery schemes by up to 5dB PSNR.