Web and Big Data. APWeb-WAIM 2023 International Workshops

The integration of Large Language Models (LLMs) and Knowledge Graphs (KGs) has emerged as a vibrant research area in the field of Natural Language Processing (NLP). However, existing approaches need help effectively harnessing the complementary strengths of LLMs and KGs. In this paper, we propose a novel system that addresses this gap by enabling bidirectional conversion between LLMs and KGs. We leverage external knowledge to enhance LLMs for domain-specific responses and fine-tune LLMs for information extraction to construct the Knowledge Graph. Moreover, users can interact with the KG, initiating new rounds of questioning in LLMs. The evaluation results highlight the effectiveness of our approach. Our system showcases the potential of combining LLMs and KGs, paving the way for advanced natural language understanding and generation in various domains.

Knowledge graphs organize entity relations using a graph structure, facilitating knowledge representation. In research, relation prediction within knowledge graphs plays a crucial role, aiding inference, latent knowledge discovery, and revealing intricate associations between entities. We present an overview of this field’s development and methods. Initially, we introduce fundamental concepts, relation prediction task definitions, and evaluation metrics. Subsequently, we delve into research, spanning rule-based, statistical, and modern approaches like representation learning, deep learning and large language models. We explore transductive and inductive learning modes, discussing their relevance in relation prediction, and classify and summarize these methods. Additionally, we evaluate method strengths, weaknesses, and suitable scenarios, providing insights. Finally, we address future research directions and challenges in knowledge graph relation prediction, offering guidance for further study and practical applications.

Reasoning commonsense knowledge is essential for Artificial Intelligence, which requires high-quality commonsense knowledge. Recently, much progress has been made in automatic commonsense knowledge generation. However, most of the works focus on obtaining positive knowledge and lack negative information. Only a few works capture the importance of negative statements, but they struggle to produce high-quality knowledge. Although some efforts have been made to generate negative statements, they fail to consider the taxonomic hierarchy between entities and are not generally applicable, leading to the generation of low-quality negative samples. To resolve the issue, we put forward Negation, a framework for effectively generating hard negative knowledge. For each entity in the commonsense knowledge base, congeners are identified with hierarchical and semantic information. Then, negative candidates are produced by replacing the entity with congeners in each triple. In order to make negative knowledge more confusing and avoid false positive examples, we design two filtering steps to remove the amount of meaningless candidates. We empirically evaluate our proposed method Negation on the downstream task, and the results demonstrate that Negation and its components effectively help generate high-quality negative knowledge.

Group recommendation can recommend satisfactory activities to group members in the recommendation system. In the research of group recommendation, the main issue is how to combine the preferences of different group members. Most of the existing group recommendations adopt a single aggregation strategy to aggregate the preferences of different group members, unable to fulfill the needs of diversified group decision-making. At the same time, most of these group recommendation methods rely on intuition or hypothesis to analyze the influence of group members, which lacks convincing theoretical support. To overcome this issue, we propose the Diversified Group Recommendation Model for social network (DGRM). This model considers two aspects of social choice and social influence, models the diversity of groups, and adopts different group recommendation strategies for different groups, which can better meet the diverse needs of users and groups. We propose a group recommendation strategy based on score fusion, which can better meet the diverse needs of users and groups. Firstly, a matrix factorization-based individual rating prediction method and a Bayesian model-based individual rating prediction method are proposed, respectively, to predict the individual ratings based on user-item interactions. Secondly, different strategies for scoring fusion are proposed, and group recommendations are made based on the fused scores for different types of groups. Finally, we verify the feasibility and effectiveness of the key technologies proposed in this paper by conducting experiments, which demonstrates the effectiveness of our proposed methods.

Nowadays, recommendation systems have been widely used in various aspects such as news, movies, music, videos, academia, and many more. The advent of personalized recommendation systems has significantly enhanced the efficiency of users’ access to information and also improved their overall experience. As an essential component of the recommendation system research field, interpretable recommendations emphasize the need to provide users with recommended results along with the rationale behind them. Unlike traditional recommendation systems, interpretable systems can not only enhance system transparency but also increase user trust and acceptance, the likelihood of users choosing the recommended product, and overall satisfaction. However, most existing interpretable recommendation systems depend on user similarity, item similarity, scoring data, or review a single aspect of the data to produce an interpretation, which makes it challenging to create trustworthy interpretable due to the limited use of these factors and intelligent reasoning. To solve this problem, we propose the Probabilistic Soft Logic (PSL)-based Interpretable Generation Model for Recommendation (called PIGM). Unlike traditional interpretable recommendation models, our PIGM takes into account user similarity, item similarity, and scoring data, employing PSL to model these factors and utilizing intelligent reasoning to generate interpretations. Firstly, based on matrix decomposition to predict the user’s score of the item. Secondly, the data is partitioned into observation dataset, target dataset and fact dataset. Thirdly, predicates are defined according to the data and the required results, and rules are defined according to the defined predicates, fact logic and relevant knowledge generated by recommendation system interpretation. Then, the weights of the rules are learned based on the maximum likelihood estimation. Finally, based on PSL reasoning and according to the defined rules, generate a recommendation list and corresponding interpretation for the user. The experimental results demonstrate the effectiveness of our proposed PIGM model on the real dataset.

Personal credit plays a vital role in the modern economy and society. However, the traditional centralized credit model suffers from numerous issues, including privacy breaches, data misuse, and unclear data ownership. Moreover, this model lacks an efficient sharing mechanism, resulting in data dispersion and low utilization rates. To solve these problems, this paper proposes a personal credit data security sharing scheme based on multi-chain collaboration. In this scheme, credit data is stored in IPFS, and data summaries and information are uploaded to the master chain to ensure data integrity and consistency. Additionally, access control and key management operations are transferred from the master chain to reduce pressure and enable capacity expansion. Furthermore, to safeguard the security and privacy of data sharing, the paper designs a more efficient access control model called PT-ABAC, based on attribute-based access control (ABAC) and capability-based access control (CapBAC). The verification of static attributes is replaced by granting user permission token, and the zero-knowledge proof algorithm is used to protect the privacy of user attributes. The model realizes the fine-grained access control of data, and effectively improves the efficiency of data access. The analysis shows that the credit data sharing scheme in this paper is characterized by distributed, tamper-proof, traceable, secure and transparent data sharing, and fine-grained access control, which ensures the authenticity and immutability of the credit data, implements strict access control in credit data sharing, and clarifies the ownership and control of the credit data.

Owing to the open and sharing characteristics, blockchain can be applied for data ownership management in data circulation. The smart contract, as a kernel technique of blockchain, is a program code that can automatically execute the business process of the scene. Smart contracts require programmers with professional coding ability in contract design and implementation. Hence, it is necessary to design an auxiliary tool for non-coding personnel to write smart contracts and manage data ownership as well. To cope with this problem, this paper proposes an advanced smart contract language OR-SPESC for data ownership. OR-SPESC is a contract language similar to natural language including five parts of parties, data assets, deeds, terms, and contract properties. It can select the appropriate function meta-language for a specific scenario to complete the scenario business. Firstly, OR-SPESC makes a formal definition of deed, which describes the ownership between parties and data assets. This paper also proposes three operations on data ownership including creation, destruction, and transfer. Then, taking the trade scenario as an example, the design of OR-SPESC contract rules. Finally, the experiments implement the tracking of OR-SPESC contract conversion and the analysis and verification of the conversion rate (CR) and product rate (PR) of OR-SPESC contract conversion. The research illustrates the effectiveness of OR-SPESC contract design and the efficiency of contract conversion.