Reporting by XinZhiyuan
Source: Specialized
【Introduction to New Zhiyuan】The event knowledge map is the focus of attention at present. But what exactly is an event knowledge graph? The Institute of Computing of the Chinese Academy of Sciences published the review paper "Event Knowledge Graph Review", which provides a comprehensive review of EKG from the aspects of history, ontology, example and application view.
In addition to entity-centric knowledge (often organized in the form of a knowledge graph (KG),events are also an essential type of knowledge in the world, triggering the rise of event-centric representations of knowledge (EKG). It plays an increasingly important role in many machine learning and ARTIFICIAL applications, such as intelligent search, question answering, recommendations, and text generation.
This paper provides a comprehensive review of EKG from the aspects of history, ontology, examples, and application views. To describe the EKG more fully, we will focus on its history, definition, pattern generalization, acquisition, and related representative graphics/systems and applications. Its development process and trends were studied. We further summarize the development direction of future EKG research.
https://arxiv.org/abs/2112.15280
The Knowledge Graph (KG) is a popular representation of knowledge released by Google in 2012. It focuses on nominal entities and their relationships and therefore represents static knowledge. However, there is a wealth of event information in the world that conveys dynamic procedural knowledge. Therefore, event-centric representations of knowledge, such as Event KG (EKG), are also important, combining entities and events. It facilitates many downstream applications such as smart search, Q&A, recommendation, and text generation[1], [2], [3], [4], [5].
This article discusses the concept of EKG and its development in depth. What do you want to know about EKG? You may be interested in its generation, the so-called EKG, how to build it, and its further applications. To provide a comprehensive introduction to EKG, we have introduced it in terms of history, ontology, examples, and application views. From a historical point of view, we introduce a brief history of EKG and our derived definition of EKG.
From the perspective of ontology, the basic concepts related to EKG, as well as EKG-related tasks and methods, are proposed, including event mode induction, script induction and EKG mode induction.
From the Instance View, we elaborate on event acquisition and EKG-related representation diagrams/systems. Specifically, the focus of event acquisition is on how to build a basic EKG and get a better EKG. The former includes event extraction and event relationship extraction, which is the most basic task. The latter includes event cross-reference resolution and event parameter completion.
From the perspective of application, some basic applications are introduced, including script event prediction and time KG prediction, as well as some deep applications such as search, Q&A, recommendation and text generation. The development process and trend of related tasks are also studied and analyzed in depth. Then point out the direction of the future.
History of the Event Knowledge Graph
EKG related concepts
What is EKG? Historical perspective
In this section, we will briefly introduce the history of EKG from a historical perspective. We then derive the definition of EKG based on historically associated concepts.
What is EKG? Ontological perspective
From an ontological perspective, we looked at patterns and related tasks. The EKG's pattern describes the basic concepts that make it up, such as the type of event, the role of event parameters, and the relationships between events. The roles of event types and event parameters make up the framework for events, the event pattern. For relationships between events, a typical script[30] organizes a set of events based on a number of event relationships that together describe common scenarios.
What is EKG? Example perspective
From the instance view, this section describes how to build an EKG, that is, a representative diagram/system related to event acquisition and EKG.
Event extraction
Event relationship extraction
Future directions and challenges
There are many research and achievements on EKG. However, there are still several directions that need attention and further study. In this section, we'll dive into these future directions.
High-performance event acquisition
Recent event acquisition studies are far from adequate in terms of effectiveness and efficiency. In particular, the accuracy of event extraction and event relationship extraction is low. This hinders the construction of high-quality basic EKG. In addition, existing models often do not focus on complexity issues. However, models with high parameter complexity and high temporal complexity are not conducive to rapidly building EKG from large amounts of data. Therefore, efficient event acquisition is an important direction for the future.
Multimodal knowledge processing
In the real world, events may be presented as text, images, audio, and video. However, existing research on EKG focuses on text processing, while ignoring a lot of information in images, audio, and video. There is little research on multimodal event representation learning [214] and event extraction [215]. In fact, events of different modes can be disambiguated and complement each other. Therefore, the combined utilization of multimodal information is an important direction in the future. Specifically, events from all modes should be represented in a unified framework, event acquisition studies should pay attention to multimodal extraction, and EKG graph inference should also consider multimodal information.
Interpretable EKG studies
In the EKG study, the study focused on fitting training data using deep learning methods. However, they often lack explainability, that is, there is no clear idea of why and how they work. In fact, knowing the reasons for the end result helps to adopt them in practical applications. It is a friendly and convincing explanation of why the end result is given. Future research on interpretable EKG will be an important direction.
Practical EKG research
Among the EKG-related tasks, some tasks are too idealistic in form and far from the real scene. For example, in an existing event, only one missing parameter or parameter role is completed, a future script event is predicted by selecting it from several candidates, and only one element of the future event is predicted. Conducting research in a more practical form is more challenging, but also more interesting, with significant implications for application.
EKG is important for many, including smart search, question answering, recommendations, and text generation. This paper reviews the EKG study from different perspectives. In particular, we delved into the history, ontology, examples, and application views of EKG. Its history, definitions, pattern induction, acquisition, related representation diagrams/systems, and applications have been studied in depth. According to its development trend, the prospect direction of future EKG research is further summarized.