Time: 08:45-10:15, October 12, 2023

Title: Knowledge Analysis, Extraction and Enhancement in Pre-trained Language Models

Abstract : Recently, large-scale pre-trained language models have made remarkable progress in knowledge-intensive natural language processing tasks. It seems to indicate that pretrained language models can naturally learn extensive knowledge from the corpus and implicitly encode it in the parameters. However, the underlying mechanisms behind the phenomenon remain unknown. Questions such as what knowledge has been acquired by language models, how to extract and utilize the knowledge, and how external knowledge can be incorporated to address the limitations of models, are all awaiting further exploration. In this tutorial, we will focus on introducing recent research advancements in the knowledge analysis, knowledge extraction, and knowledge enhancement of pre-trained language models.

Speaker: Yubo Chen is an Associate Researcher at the Institute of Automation, Chinese Academy of Sciences. His research interests include Knowledge Graph, Natural Language Processing and Large Language Model. He has published 40+ papers on ACL, EMNLP, COLING, CIKM, WWW and AAAI. His work has been cited over 4500 times on Google Scholar. Two of his papers have been selected as high-impact papers at ACL and EMNLP (Paper Digest selection), and he has received multiple Best Paper Awards (NLP-NABD 2016, CCKS 2017, CCL 2020, CCKS 2020). He was selected for the 5th China Association for Science and Technology Youth Talent Lifting Project in 2020, and was recognized as a Global Chinese AI Young Scholar in 2022, a member of the Youth Innovation Promotion Association of the Chinese Academy of Sciences in 2022. He serves as the Secretary-General of the Youth Working Committee of the Chinese Information Processing Society of China, Area Chair of the COLING 2022, Editorial Board Member of Data Intelligence. He was awarded the first prize of the "Qian Weichang Chinese Information Processing Science and Technology Award" by the Chinese Information Processing Society of China in 2018 and the first prize of the Beijing Science and Technology Progress Award in 2019.

Time: 10:30-12:00, October 12, 2023

Title: Recent Advances in LLM-based Autonomous AI Agent

Abstract : In recent years, research on AI agents based on large language models, represented by Generative Agents, has attracted widespread attention from researchers. This report focuses on the construction of autonomous AI agents based on large language models and their applications in the field of user behavior analysis. It provides an overview of key technologies for building general and domain-specific AI agents using large language models, analyzes the existing challenges in this field, and outlines future development directions.

Speaker: Xu Chen is a tenure-track associate professor from Gaoling School of Artifitial Intellengence, Renmin University of China. Before joining Renmin University of China, he was a research fellow at University College London, UK. Xu Chen obtained his PhD degree from Tsinghua University. His research interests lie in recommender system, causal inference and reinforcement learning. He has published more than 60 papers on top-tier conferences/journals like TheWebConf、AIJ、TKDE、SIGIR、WSDM and TOIS. As co-leaders, He has develped recommender libarary "RecBole", explainable recommendation dataset "REASONER" and LLM-based recommender simulator "RecAgent". His papers have won the best resource paper runner up award on CIKM 2022, best paper honorable mention award on the Web Conference 2018 and best paper award on AIRS 2017. He has also obtained CCF Science and Technology Achievement Award (Natural Science Award Second Prize), and ACM-Beijing Rising Star award.

Time: 14:00-15:30, October 12, 2023

Title: Towards Deep, Broad, and Dynamic Language Understanding on Social Media

Abstract : Social media platforms have become the popular outlet for people to share ideas, voice opinions, and discuss topics they are interested in. Meanwhile, the deluge of overloaded content streaming through social media in real-time is far outpacing the human capacity to read and understand language. It consequently presents a pressing need to automate the process of natural language understanding (NLU) on social media. Today, although the state-of-the-art (SOTA) NLU models based on large language modes (LLMs) have demonstrated promising results on many benchmarks, the realistic social media scenarios may still exhibit grand challenges beyond the capabilities of SOTA NLU models to effectively tackle. On the one hand, models would face ambiguous language, users in diversity, and context ever evolving, all requiring models to cope with deep semantics, broad users, and dynamic environments. On the other hand, existing NLU models, far from the mastery of language grounding in the world, are unable to well deduce meanings of the unknown and ideas between lines.
In light of these concerns, this talk will center around a key question: what do we essentially need to apply NLU in the real social media world? I will summarize my past, present, and future research of how to enable NLU models to better tackle deep semantics in ambiguous expressions (deep understanding), broad varieties of social media users (broad understanding), and dynamic environments continuously generating new features (dynamic understanding). From here we will start a journey towards a more robust NLU on social media, which will allow the real-world applications to be genuinely benefited from the innovative research findings.

Speaker: Dr. Jing Li is an Assistant Professor of the Department of Computing, The Hong Kong Polytechnic University (PolyU) since 2019. Before joining PolyU, she worked in the Natural Language Processing Center, Tencent AI Lab as a senior researcher from 2017 to 2019. Jing obtained her PhD degree from the Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong in 2017. Before that, she received her B.S. degree from Department of Machine Intelligence, Peking University in 2013. Jing has been working with NLP research for over 10 years, with the background from both academia and industry. She has broad research interests in Natural Language Processing, Computational Social Science, and Machine Learning. Particularly, she works on novel algorithms for language representation learning, social media language understanding, conversation and social interaction modeling, and robust NLP and multimodal applications in noisy real-world applications. She regularly publishes in the top-tier NLP conferences and journals and has been invited to serve in the organization and program committee in many of these venues. Her research is substantially supported by competitive research funds from NSFC, HKSAR Research Grant Committee (RGC), HKSAR Innovation and Technology Fund (ITF), CCF-Tencent, CCF-Baidu, etc.

Time: 15:45-17:15, October 12, 2023

Title: Medical large language models, the practice from HuatuoGPT

Abstract : The large models have brought significant changes to natural language processing. They can provide users with useful information in some general areas, enhancing learning and work efficiency. However, their performance in specialized domains still lags behind domain experts, such as experienced doctors and lawyers. To build expert-level models in specialized domains, a myriad of issues emerge, spanning evaluation, training, data, regulation, efficiency, etc. This report will focus on the medical field's Huatuo GPT as a foundation to elaborate on the challenges, practices, and future of specialized domain large models. The access link to Huatuo GPT is https://www.huatuogpt.cn/. The practices of Huatuo GPT may also offer valuable insights for other specialized domains.

Speaker: Wang Benyou is an assistant professor and a deputy director of the Health informatics Center, at the School of Data Science at the Chinese University of Hong Kong (Shenzhen). He is also a research scientist at the Shenzhen Institute of Big Data. To date, he has received the Best Paper Award honorable mention at SIGIR 2017, the Best Interpretable NLP Paper at NAACL 2019, the Best Paper Award at NLPCC2022, the Huawei Spark Award, and the European Union Marie Curie Scholarship. The research team he leads has developed large models including the multi-language Phoenix model (supporting Chinese) and the medical and health domain model, HuatuoGPT.

Time: 08:45-10:15, October 13, 2023

Title: Large Language Model-Powered Multi-Agent Interaction

Abstract : Recently, large Language Models (LLMs) have demonstrated human-like intelligence and revolutionized various applications in artificial intelligence, encouraging the emerging exploration of LLM-powered autonomous agents. With LLMs as the brain, the agent interacts with its environment to resolve complex goals, complemented by task planning, memory, tool usage, etc.
Inspired by the promising agent demos such as AutoGPT, there has been a surge in research focusing on multi-agent interactions, particularly in studying the behaviour of agents in a simulated game setting. Through competition and cooperation, multiple agents can autonomously improve each other in factuality, reasoning and safety.
There are still many unanswered questions and research directions on this topic. In this tutorial, we will first introduce the background and components of LLM-powered agents, then take three multi-agent interactions (i.e., generative agents simulation, competition and cooperation) as examples, review the technical development, and finally discuss the future direction of the topic.

Speaker: Baotian Hu is an associate professor at Computer Science Department of Harbin Institute of Technology, Shenzhen. He is also Deputy Secretary General of China Chinese Information Processing Society (CIPS) Medical Health and Biological Information Processing Professional Committee and Deputy Secretary General of the Artificial Intelligence Professional Committee of Shenzhen Computer Society. He received his B.A. in Information and Computing Science from the Shandong University of Science and Technology in 2010, and his M.S. and Ph.D. in Computer Science from the Harbin Institute of Technology in 2012 and 2016 respectively. He was a visiting researcher at Huawei Noah's Ark Laboratory (Hong Kong), a senior researcher at WeChat of Tencent Technology Co., LTD., and a postdoctoral researcher at the University of Massachusetts, USA. His research interests focus on LLM and its applications. He has published 50+ technical papers at prestigious international journals and conferences.. He was nominated for Outstanding Ph.D. Dissertation by the Chinese Language Information Society of China in 2018 and awarded as Outstanding Ph.D. Dissertation for the centennial anniversary of Harbin Institute of Technology. He was invited to serve as Area Chair (AC) or Senior Program Committee Member (SPC) at prestigious academic conferences including ACL2023, EMNLP2022, IJCAI2021/2023, AACL2022.

Time: 10:30-12:00, October 13, 2023

Title: Text-Guided Image and Video Editing: Progress and Challenges

Abstract : In recent years, diffusion models have demonstrated outstanding capabilities in visual generation. Beyond their prowess in image generation, these models have also captivated researchers by offering the potential to employ textual descriptions for guiding content editing in both images and videos, thereby fostering more creative and personalized visual presentation. However, the application of pretrained diffusion models to editing real images and videos in specific contexts remains confronted with challenges. One such challenge involves accurately capturing the diversity and contextual information of textual descriptions to ensure alignment between editing outcomes and the original intent. Additionally, achieving real-time editing and large-scale application stands as another issue to be addressed. This report delves into the challenges and obstacles faced by pretrained diffusion models in visual editing, while also providing a comprehensive overview of existing research accomplishments and showcasing the latest advances in this field.

Speaker: Jingjing Chen is an Associate Professor at the School of Computer Science and Technology, Fudan University. Before that, She obtained her Ph.D. degree from City University of Hong Kong in 2018, and from September 2018 to July 2019, she worked as a postdoctoral researcher at the National University of Singapore. Her primary research areas include multimedia content analysis, computer vision, and multimedia model security. Her research output includes over 70 papers published in prominent international conferences and journals such as ACM Multimedia, CVPR, ICCV, AAAI, ICMR, IEEE TIP, and IEEE TMM. Her research works have received recognition from academic organizations in China and beyond, and she has received several accolades, including the Best Student Paper Award at the ACM Multimedia in 2016, the Best Student Paper Award at the Multimedia Modeling conference in 2017, the Best Paper Award at the China Multimedia Conference in 2022, "ACM Shanghai Rising Star Award" in 2020 and the "AI 2000 Most Influential Scholar Nomination Award" in 2021. Moreover, she was also listed on the 2023 Baidu "AI Chinese Women's Young Scholars List".

Time: 14:00-15:30, October 13, 2023

Title: Pre-training Methods in Information Retrieval: Advances and Challenges

Abstract : In recent years, with the development of deep learning technology, the integration of information retrieval and natural language processing has become increasingly profound. Pre-trained models, as an important method in the field of NLP, have also been widely used in information retrieval. This tutorial focuses on the application of pre-trained models in information retrieval and systematically reviews related research in recent years. The main contents include: 1) background: introduction of basic concepts of information retrieval; 2) the application of pre-trained models in the first-stage retrieval: including the application of pre-trained models in sparse retrieval and dense retrieval; 3) the application of pre-trained models in the re-ranking stage: including representation learning methods, interactive learning methods, and model acceleration; 4) generative information retrieval: including document identifier repersentation, architecture design, and optimization strategy.

Speaker: Yixing Fan is an associate researcher from institute of computing technology, chinese academy of sciences. His primary research areas include information retrieval, natural language retrieval, etc. He has published around 40 papers in top-tier conferences and journals, including but not limited to SIGIR, WWW, CIKM, TOIS, FnTIR. He was selected for the 6th China Association for Science and Technology Youth Talent Lifting Project in 2021, as well as the Youth Innovation Promotion Association of the Chinese Academy of Sciences in 2021. He serves as the Local Organization Chair for the SIGIR-AP 2023 and the CCIR 2023, Program Committee Chair of WI-IAT 2022, Program Committee Member of SIGIR 2023, CIKM2023, SIGIR2022, CIKM2022, TheWebConf2022, WSDM2022, and so on.

Time: 15:45-17:15, October 13, 2023

Title: Generative Retrieval

Abstract : Generative retrieval is an emerging paradigm for information retrieval, which employs generative models to directly generate document identifiers for a given query. Compared to traditional sparse retrieval and dense retrieval, generative retrieval is an end-to-end approach and can better leverage the capabilities of recently advanced large language models. In this tutorial, I will introduce the progress we made so far towards generative retrieval, including identifier design, training strategy, dynamic corpora, and so on. I will conclude with a discussion of generative retrieval in the era of large language models and present some open challenges for future research.

Speaker: Pengjie Ren is a professor of IRLab at Shandong University. Before that, he was a postdoc researcher at the Informatics Institute, University of Amsterdam. His research interests fall in natural language processing, information retrieval, etc. He has published around 100 papers in top-tier conferences and journals, including but not limited to ACM SIGIR, The Web Conference (WWW), ACL, KDD, AAAI, EMNLP, ACM TOIS, IEEE TKDE, AIJ, etc. Several of his papers from CIKM and SIGIR are selected the most influential papers of the year. Parts of his research works won the outstanding doctoral dissertation of Shandong province, the CIKM best paper runner-up award, the NLPCC best student paper award, etc.