Invited Talks

Dr. Flavius Frasincar

Talk Title: Aspect-Based Sentiment Classification 

Abstract: The World Wide Web is the most popular platform for people to express their opinions about products or services. Being able to extract the sentiments associated with these opinions is useful for both companies and consumers alike. Due to the sheer amount of opinionated textual data, manual extraction of sentiment is virtually impossible that is why Sentiment Analysis (SA) aims to automatically extract sentiment. A popular subfield of SA is Aspect-Based Sentiment Analysis (ABSA), which proposes to extract the sentiments associated with the aspects of the entities of interest, providing thus detailed sentiment information. ABSA has two main tasks: Aspect Detection (AD), which aims to discover the aspects of the entities of interest, and Aspect-Based Sentiment Classification (ABSC), which aims to determine the sentiments associated with these aspects. This talk focuses on ABSC, where current approaches are classified as: knowledge-based, machine learning, and hybrid models. Due to their good performance, special attention is payed to deep learning solutions based on the transformer model and hybrid models that combine deep learning with knowledge bases. At the end of the talk, future research directions are identified for ABSC. This presentation is partially based on the article “A Survey on Aspect-Based Sentiment Classification” by Gianni Brauwers and Flavius Frasincar, which appeared in ACM Computing Surveys, volume 55, number 4, pages 65:1-65:37, 2023. 

Bio: Flavius Frasincar is an Assistant Professor in computer science at Erasmus University Rotterdam, the Netherlands. He received the Ph.D. degree in computer science, in 2005, from Eindhoven University of Technology, the Netherlands. He has published numerous papers in top conferences and journals in the areas of databases, Web information systems, personalization, machine learning, computational linguistics, and the Semantic Web. Two of his papers received awards: Distinguished Paper for the 16th International Conference on Web Engineering (ICWE 2016) and Best Paper for the Data & Knowledge Engineering Journal 2013. Regarding professional services, he served as the general chair of the 15th International Conference on Web Engineering (ICWE 2015), vice-general chair of the 21st International Conference on Web Engineering (ICWE 2021), and PC co-chair of the 19th International Conference on Web Engineering (ICWE 2019) and 22nd International Conference on Natural Language & Information Systems (NLDB 2017). He is a member of the editorial boards of Information Processing & Management, International Journal of Web Engineering and Technology, and Computational Linguistics in the Netherlands Journal, and co-editor-in-chief of Journal of Web Engineering. He is a member of the Association for Computing Machinery (ACM) and International Society for Web Engineering (ISWE).

Dr. Diyi Yang

Talk Title: Dynamic Testing and Efficient Unlearning in Large Language Models  

Abstract: Large language models (LLMs) have demonstrated strong performances in various tasks through pre-training on massive textual data. Such advances also come with critical challenges, including potential data contamination and privacy concerns.  We share two recent studies on how to enhance model evaluation and protect data privacy. The first part presents a novel evaluation protocol that generates evaluation samples dynamically by using directed acyclic graphs, which can be used to assess LLMs with varying degrees of complexity. The second part integrates lightweight unlearning layers into LLMs in order to enable selective forgetting without retraining the whole model. These two works highlight the need for dynamic and adaptive model evaluation and data management strategies toward more robust and responsible AI development. 

Bio: I am an assistant professor in the Computer Science Department at Stanford, affiliated with the Stanford NLP Group, Stanford HCI Group, Stanford AI Lab (SAIL), and Stanford Human-Centered Artificial Intelligence (HAI). I am interested in Socially Aware Natural Language Processing. My research goal is to better understand human communication in social context and build socially aware language technologies to support human-human and human-computer interaction. 

Dr. James Zhang

Talk Title: An Intricate Journey: Tackling Challenges in Introducing Large Language Models to Financial Applications

Abstract: 

The integration of large language model (LLM) into the financial sector necessitates intricate system engineering. In this presentation, we will outline the obstacles and complexities encountered in the real-world deployment of Ant Group's LLM in financial applications, along with the introduction to our accomplished solutions for end users.

The stringent regulatory compliance requirements of financial industry implicate that "out-of-the-box" LLMs cannot be readily employed without addressing their deficiencies, which is further complicated by practical challenges, such as LLMs’ lack of domain knowledge and incompetence in complex decision-making, as well as their reliability issues.

Another obstacle is the the ‘last-mile’ of LLM applications, which necessitates the following ad hoc solutions: 1) the amalgamation of ad hoc enhancement technologies, i.e., the harmonization of general-purpose LLM and ad hoc small models.  2) The integration of knowledge augmentation technology to complement LLM’s hallucination. 3)The fusion of retrieval enhancement technologies combining both Closed QA and Open QA.

Reliability is another major issue of LLM applications.  Even though the aforementioned combinations of technologies can effectively compensate for the limitations of LLMs, they must achieve even higher reliability before being introduced to financial consumers. To ensure financial reliability of LLM services, we proposed the following three principles: (1) LLM applications should exercise caution in financial decision-making, clearly delineating capability boundaries, business boundaries, and compliance boundaries; (2) The values of the LLM’s output must align with the values of regulating authorities; (3) Core indicators such as security, compliance, accuracy, and latency must meet rigorous industry standards.  In order to achieve these goals, we implemented the following procedures: First, prioritizing quality over quantity in data sources, and we invested heavily on both manpower and automation to collect, clean and select data for training, finetuning and alignment. Second, during the online reasoning stage, we implemented a dedicated security barrier for financial LLMs, i.e., strict risk identification and interception are conducted at both the request and generation interfaces.   This system, known as "AntScreen" comprises hundreds of recognition models and hundreds of thousands of discrimination rules.  Simultaneously, millions of negative samples have been assembled to continuously and automatically probe potential reliability weaknesses of LLMs and to assess security capabilities.  Through the reinforcement of a series of security and compliance technologies, our LLM has achieved over 99% in various dimensions of financial reliability assessment, comparable to human experts, and therefore meets the reliability requirements for consumers.

Bio: Dr. James Zhang is the Managing Director of AI Forecasting and Strategy Platform of Ant Group. Dr. Zhang obtained his Ph.D. degree from Univ. of Ottawa, Canada in Electrical Engineering, and both his Master's and Bachelor degrees from Zhejiang Univ., China. Before he joined Ant Financial, Dr. Zhang worked on finance-related AI at Bloomberg, helped setting up the AI branch of Bloomberg Labs and initiating the GPU computation farm of Bloomberg. Dr. Zhang worked in various areas including image processing, natural language processing, pattern recognition, high-speed hardware development, optical networks, operations research, biometrics, and financial systems.