智慧型語言演算法實驗室

@article{10.1145/3744341, author = {Yung, Yiu Cheong and Lin, Ying-Jia and Kao, Hung-Yu}, title = {Exploring the Effectiveness of Pre-training Language Models with Incorporation of Diglossia for Hong Kong Content}, year = {2025}, issue_date = {July 2025}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, volume = {24}, number = {7}, issn = {2375-4699}, url = {https://doi.org/10.1145/3744341}, doi = {10.1145/3744341}, abstract = {In this article, we present our works to create the first Hong Kong content-based public pre-training dataset and the experiments which resulted in the creation of ELECTRA-based models for commonly used languages in Hong Kong. The creation of pre-training dataset is required for us to study the effect of diglossia on Hong Kong language model, and this is the first ever study on the effect starting all the way from dataset creation phase. Our experiment shows that removing diglossia from pre-training data hurts model performance. We will release our data and models to encourage future studies in Hong Kong languages.1}, journal = {ACM Trans. Asian Low-Resour. Lang. Inf. Process.}, month = jul, articleno = {71}, numpages = {16}, keywords = {Hong Kong, diglossia, ELECTRA, language modeling} }

@Article{s23031483, AUTHOR = {Lin, Ying-Jia and Chen, Kuan-Yu and Kao, Hung-Yu}, TITLE = {LAD: Layer-Wise Adaptive Distillation for BERT Model Compression}, JOURNAL = {Sensors}, VOLUME = {23}, YEAR = {2023}, NUMBER = {3}, ARTICLE-NUMBER = {1483}, URL = {https://www.mdpi.com/1424-8220/23/3/1483}, PubMedID = {36772523}, ISSN = {1424-8220}, ABSTRACT = {Recent advances with large-scale pre-trained language models (e.g., BERT) have brought significant potential to natural language processing. However, the large model size hinders their use in IoT and edge devices. Several studies have utilized task-specific knowledge distillation to compress the pre-trained language models. However, to reduce the number of layers in a large model, a sound strategy for distilling knowledge to a student model with fewer layers than the teacher model is lacking. In this work, we present Layer-wise Adaptive Distillation (LAD), a task-specific distillation framework that can be used to reduce the model size of BERT. We design an iterative aggregation mechanism with multiple gate blocks in LAD to adaptively distill layer-wise internal knowledge from the teacher model to the student model. The proposed method enables an effective knowledge transfer process for a student model, without skipping any teacher layers. The experimental results show that both the six-layer and four-layer LAD student models outperform previous task-specific distillation approaches during GLUE tasks.}, DOI = {10.3390/s23031483} }

@article{AL:10162364-N202403130009-00014, title ={Relation-Aware Image Captioning with Hybrid-Attention for Explainable Visual Question Answering}, author ={YING-JIA LIN and CHING-SHAN TSENG and HUNG-YU KAO}, keywords ={visual question answering; explainable VQA; multi-task learning; graph attention networks; vision-language model}, journal ={Journal of Information Science and Engineering}, volume ={40}, number ={3}, year ={2024}, month ={May}, abstract ={Recent studies leveraging object detection as the preliminary step for Visual Question Answering (VQA) ignore the relationships between different objects inside an image based on the textual question. In addition, the previous VQA models work like black-box functions, which means it is difficult to explain why a model provides such answers to the corresponding inputs. To address the issues above, we propose a new model structure to strengthen the representations for different objects and provide explainability for the VQA task. We construct a relation graph to capture the relative positions between region pairs and then create relation-aware visual features with a relation encoder based on graph attention networks. To make the final VQA predictions explainable, we introduce a multi-task learning framework with an additional explanation generator to help our model produce reasonable explanations. Simultaneously, the generated explanations are incorporated with the visual features using a novel Hybrid-Attention mechanism to enhance cross-modal understanding. Experiments show that the proposed method performs better on the VQA task than the several baselines. In addition, incorporation with the explanation generator can provide reasonable explanations along with the predicted answers.}, pages ={649-659}, ISSN ={1016-2364}, DOI ={10.6688/JISE.202405_40(3).0014}, }

@inproceedings{yang-etal-2025-maple, title = "{MAPLE}: Enhancing Review Generation with Multi-Aspect Prompt {LE}arning in Explainable Recommendation", author = "Yang, Ching-Wen and Feng, Zhi-Quan and Lin, Ying-Jia and Chen, Che Wei and Wu, Kun-da and Xu, Hao and Jui-Feng, Yao and Kao, Hung-Yu", editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher", booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", month = jul, year = "2025", address = "Vienna, Austria", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2025.acl-long.1535/", doi = "10.18653/v1/2025.acl-long.1535", pages = "31803--31821", ISBN = "979-8-89176-251-0", abstract = "Explainable Recommendation task is designed to receive a pair of user and item and output explanations to justify why an item is recommended to a user. Many models approach review generation as a proxy for explainable recommendations. While these models can produce fluent and grammatically correct sentences, they often lack preciseness and fail to provide personalized informative recommendations. To address this issue, we propose a personalized, aspect-controlled model called Multi-Aspect Prompt LEarner (MAPLE), which integrates aspect category as another input dimension to facilitate memorizing fine-grained aspect terms. Experiments conducted on two real-world review datasets in the restaurant domain demonstrate that MAPLE significantly outperforms baseline review-generation models. MAPLE excels in both text and feature diversity, ensuring that the generated content covers a wide range of aspects. Additionally, MAPLE delivers good generation quality while maintaining strong coherence and factual relevance. The code and dataset used in this paper can be found at https://github.com/Nana2929/MAPLE." }

@InProceedings{10.1007/978-981-96-8173-0_15, author="Hsu, Chih-Hao and Lin, Ying-Jia and Kao, Hung-Yu", editor="Wu, Xintao and Spiliopoulou, Myra and Wang, Can and Kumar, Vipin and Cao, Longbing and Wu, Yanqiu and Yao, Yu and Wu, Zhangkai", title="From Persona to Person: Enhancing the Naturalness with Multiple Discourse Relations Graph Learning in Personalized Dialogue Generation", booktitle="Advances in Knowledge Discovery and Data Mining", year="2025", publisher="Springer Nature Singapore", address="Singapore", pages="187--198", abstract="In dialogue generation, the naturalness of responses is crucial for effective human-machine interaction. Personalized response generation poses even greater challenges, as the responses must remain coherent and consistent with the user's personal traits or persona descriptions. We propose MUDI (Multiple Discourse Relations Graph Learning) for personalized dialogue generation. We utilize a Large Language Model to assist in annotating discourse relations and to transform dialogue data into structured dialogue graphs. Our graph encoder, the proposed DialogueGAT model, then captures implicit discourse relations within this structure, along with persona descriptions. During the personalized response generation phase, novel coherence-aware attention strategies are implemented to enhance the decoder's consideration of discourse relations. Our experiments demonstrate significant improvements in the quality of personalized responses, thus resembling human-like dialogue exchanges.", isbn="978-981-96-8173-0" }

@article{Lin_Lin_Yeh_Li_Hu_Hsu_Lee_Kao_2024, title={CFEVER: A Chinese Fact Extraction and VERification Dataset}, volume={38}, url={https://ojs.aaai.org/index.php/AAAI/article/view/29825}, DOI={10.1609/aaai.v38i17.29825}, abstractNote={We present CFEVER, a Chinese dataset designed for Fact Extraction and VERification. CFEVER comprises 30,012 manually created claims based on content in Chinese Wikipedia. Each claim in CFEVER is labeled as "Supports", "Refutes", or "Not Enough Info" to depict its degree of factualness. Similar to the FEVER dataset, claims in the "Supports" and "Refutes" categories are also annotated with corresponding evidence sentences sourced from single or multiple pages in Chinese Wikipedia. Our labeled dataset holds a Fleiss' kappa value of 0.7934 for five-way inter-annotator agreement. In addition, through the experiments with the state-of-the-art approaches developed on the FEVER dataset and a simple baseline for CFEVER, we demonstrate that our dataset is a new rigorous benchmark for factual extraction and verification, which can be further used for developing automated systems to alleviate human fact-checking efforts. CFEVER is available at https://ikmlab.github.io/CFEVER.}, number={17}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, author={Lin, Ying-Jia and Lin, Chun-Yi and Yeh, Chia-Jen and Li, Yi-Ting and Hu, Yun-Yu and Hsu, Chih-Hao and Lee, Mei-Feng and Kao, Hung-Yu}, year={2024}, month={Mar.}, pages={18626-18634} }