Knowledge Graphs: Opportunities and Challenges - Leitura de Artigo

Peng, Ciyuan & Xia, Feng & Naseriparsa, Mehdi & Osborne, Francesco. (2023). Knowledge Graphs: Opportunities and Challenges. Artificial Intelligence Review. 1-32. 10.1007/s10462-023-10465-9. 

 

2.2.2 Knowledge Acquisition

[Extrair e Delimitar Contexto?]  

2.2.3 Knowledge Graph Completion

[Completar Contexto?]

 

2.2.4 Knowledge Fusion

[Alinhamento de Entidades e as questões da Identidade e resolução de conflitos.]

 

2.2.5 Knowledge Reasoning

[Completar Contexto? Detectar controversas]

 

2.2.6 AI Systems

[Usar KG para contextualizar perguntas e respostas]

 

3.2 Question–Answering Systems

There are two main types of questions in this space: simple and multi-hop questions, respectively. Simple questions are answered only by referring to a single triplet, while multi-hop questions require combining multiple entities and relations. Focusing on simple questions, Huang et al. (2019) proposed a knowledge graph embedding-based question-answering system (KEQA). They translated the question and its corresponding answer into a single triplet. For instance, the question “ Which film acted by Leonardo" and one of its answers “Inception" can be expressed as the following triplet: (Leonard, act, Inception).

 

3.3 Information Retrieval

Compared to traditional information retrieval, the knowledge graph-based information retrieval has the following advantages:
• Semantic Representation of Items: Items are represented according to a formal and interlinked model that supports semantic similarity, reasoning, and query expansion. This typically allows the system to retrieve more relevant items and makes the system more interpretable.
• High Search Efficiency: Knowledge graph-based information retrieval can use the advanced representation of the items to reduce the search space significantly (e.g., discarding documents that use the same terms with different meanings), resulting in improved efficiency.
• Accurate Retrieval Results: In knowledge graph-based information retrieval, the correlation between query and documents is analyzed based on the relations between entities in the knowledge graph. This is more accurate than finding the similarities between queries and documents.      

 

4 Applications and Potentials

 

4.2 Scientific Research

A variety of knowledge graphs focus on supporting the scientific process and assisting researchers in exploring research knowledge and identifying relevant materials (Xia et al. 2016). They typically describe documents (e.g., research articles, patents), actors (e.g., authors, organizations), entities (e.g., topics, tasks, technologies), and other contextual information (e.g., projects, funding) in an interlinked manner. For instance, Microsoft Academic Graph (MAG) (Wang et al. 2020a) is a heterogeneous knowledge graph. MAG contains the metadata of more than 248M scientific publications, including citations, authors, institutions, journals, conferences, and fields of study. The AMiner Graph (Zhang et al. 2018) is the corpus of more than 200M publications generated and used by the AMiner system1 . The Open Academic Graph (OAG)2 is a massive knowledge graph that integrates Microsoft Academic Graph and AMiner Graph. AceKG (Wang et al. 2018c) is a large-scale knowledge graph that provides 3 billion triples of academic facts about papers, authors, fields of study, venues, and institutes, as well as the relations among them.

 

[Outros exemplos de KG abertos que pode ter o contexto mapeado]

4.3 Social Networks

One of the biggest problems in this space is fake news (Zhang et al. 2019a). Online social media has become the principal platform for people to consume news. Therefore, a considerable amount of research has been done for fake news detection (Choi et al. 2020; Meel and Vishwakarma 2020). Most recently, Mayank et al. (2021) exploited a knowledge graph-based model called DEAP-FAKED to detect fake news on social media. Specifically, DEAP-FAKED learns news content and identifies existing entities in the news as the nodes of the knowledge graph. Afterward, a GNN-based technique is applied to encode the entities and detect anomalies that may be linked with fake news.

5 Technical Challenges

5.2 Knowledge Acquisition

In addition, a domain-specific knowledge graph schema is knowledge-oriented, while a constructed knowledge graph schema is data-oriented for covering all data features (Zhou et al. 2022). Therefore, it is inefficient to produce domain-specific knowledge graphs by extracting entities and properties from raw data. Hence, it is an essential issue to efficiently achieve knowledge acquisition tasks by generating domain-specific knowledge graphs.

5.3 Knowledge Graph Completion

Knowledge graph completion methods assume knowledge graphs are static and fail to capture the dynamic evolution of knowledge graphs. To obtain accurate facts over time, temporal knowledge graph completion, which considers the temporal information reflecting the validity of knowledge, has emerged. Compared to static knowledge graph completion, temporal knowledge graph completion methods integrate timestamps into the learning process. Hence, they explore the time-sensitive facts and improve the link prediction accuracy significantly. Although temporal knowledge graph completion methods have shown brilliant performance, they still face serious challenges. Because these models consider time information would be less efficient (Shao et al. 2022), the key challenge of temporal knowledge graph completion is how to effectively incorporate timestamps of facts into the learning models and properly capture the temporal dynamics of facts.

[Contexto Temporal]

5.5 Knowledge Reasoning

There are two tasks in knowledge reasoning: single-hop prediction and multi-hop reasoning (Ren et al. 2022). Single-hop prediction predicts one element of a triplet for the given two elements, while multi-hop reasoning predicts one or more elements in a multi-hop logical query. In other words, in the multi-hop reasoning scenario, finding the answer to a typical question and forming new triplets requires the prediction and imputation of multiple edges and nodes. Multi-hop reasoning achieves a more precise formation of triplets when compared with the single-hop prediction. Therefore, multi-hop reasoning has attracted more attention and become a critical need for the development of knowledge graphs in recent years. Although many works have been done, multi-hop reasoning over knowledge graphs remains largely unexplored. Notably, multi-hop reasoning on massive knowledge graphs is one of the challenging tasks (Zhu et al. 2022).