Event prediction based on evolutionary event ontology knowledge

Highlights

• We build EEOK that leverages the evolutionary event knowledge represented in standard ontology language OWL for representing a set of evolutionary patterns.

• We propose a framework with a pipeline procedure from event extraction to event prediction.

• Considering the different event domains, we offer a domain-aware event prediction method which has been shown superiority over existing approaches.

Abstract

The evolution and development of breaking news events usually present regular patterns, leading to the happening of sequential events. Therefore, the analysis of such evolutionary patterns among events and prediction to breaking news events from free text is a valuable capability for decision support systems. Traditional systems tend to focus on contents distribution information but ignore the inherent regularity of evolutionary events. We introduce evolutionary event ontology knowledge (EEOK) structuring the evolutionary patterns in five different event domains, namely Explosion, Conflagration, Geological Hazard, Traffic Accident, Personal Injury. Based on EEOK which provides a representing general-purpose ontology knowledge, we also explore a framework with a pipeline semantic analysis procedure of event extraction, evolutionary event recognition, and event prediction. Since the evolutionary event under each event domain has different evolution patterns, our proposed event prediction model combines the event types to capture the inherent regulation of evolutionary events. Comparative analyses are presented to show the effectiveness of the proposed prediction model compared to other alternative methods.

Introduction

Social media platforms like Twitter, Facebook, and Weibo generate a large amount of online news to discover, report, share, and communicate with others about various public events. The burst of netizens’ activities on the web can be seen as a valuable real-time reflection of events as they happen in real-time. Modeling evolutionary events and predicting subsequent events is crucial to many applications for text mining. Recent developments in the field of textual inferences and analytics have led to a renewed interest in event prediction. It is becoming an increasingly crucial issue in a broad range of fields, including finance, security, policy/governance, NGO planning, and disaster coordination efforts [1], [2]. Some related works focus on predicting news event’s importance and impact [3], [4] based on the keywords volume approach [5], [6], and sequential clustering [7]. Other prevailing event prediction methods focus on the dynamic content on the web with chronological sequence or spatial location of event statements, including Time-series (TS) methods [8], [9], [10] and Spatio-temporal (ST) methods [11], [12]. Several works focus on tracking social media for event prediction on the stock market [13], [14], traffic prediction [15], [16], spatial crime analysis [17] and malware propagation [18], [19]. Some researches have modeled the latent factor of user activities history to exploit the timestamped interaction information among users for predicting real-time events [20] simultaneously.

Events are presented in online news via the indication of anchor/trigger words reflecting that something has happened. These main words evoking the events, are called event mentions [21]. To make the event mentions be useful (i.e., for extracting knowledge and event prediction), some current works have also been proposed based on leveraging knowledge represented from unstructured texts for events prediction. The unstructured text contains contextual explanations for many of event knowledge and evolutionary patterns that are correlative to news events. Recent pieces of literatures offer findings of explicit or implicit causal relations between two events to construct causal patterns among events. As shown in Fig. 1, this is an event evolutionary pattern which the seed event is ‘a fire outbreak in somewhere’. Actually, each event is a generalized schema which performs integration of event instances. For example, an event instance of ‘cause[fire, 2 people are badly burnt]’ and another instance of ‘cause[fire, 3 people died]’ can be mapped into a event knowledge of ‘cause[fire, injuries and deaths]’. Then, subsequent events catenate by relations one-by-one to form an evolutionary pattern.

These evolutionary event patterns extracted from event texts are used to predict special events. Esteban et al. [22] maps a knowledge pattern graph to a tensor representation for event prediction. The method of using causal patterns is called textual semantic events prediction (TSEP), which stands out in many other techniques like TS or ST et al. Given the occurred events, TSEP aims to predict the plausible subsequent events with knowledge inference. For example, it is observed that when an earthquake occurs in some areas, plagues also break out. So as an earthquake occurs in another area, considering this area’s environment and post-disaster measures are similar to that of the former, we can expect an outbreak of the plagues in this area.

However, to generate a plausible subsequent event for prediction, there exists difficulties and challenges which are the following: (1) Lack of dynamic extensibility to extract and predict the event. Existing TSEP methods construct patterns of evolutionary events through lexical granularity for event prediction. These methods are limited to source text cannot be automatically extended with knowledge. (2) Lack of ontology knowledge of the evolutionary event. Some specific patterns exist in the evolutionary event, and the ontology of an evolutionary event can support the modeling of such patterns. (3) Lack of appropriate computational knowledge-augmented models. Modeling of event prediction requires computational frameworks for receiving extensive source data for event extraction. It also requires integrating evolutionary event patterns.

Our work stands out in many aspects from those mentioned methods like TS, ST, and TSEP. We introduce a semantic and language-independent system to predict news events by logical knowledge unrestricted specific textual scenarios. Based on informative news texts from Baidu encyclopedia entries, we extract a large number of evolutionary event patterns and represent them in standard web ontology language (OWL). The events extracted from entries are transformed into OWL structure and merged into the knowledge base. OWL is inherently useful for retrieving, reasoning, and eventually prediction by the machine. These event knowledge with OWL formulation are sufficient to understand a text with logical rules or patterns semantically. It is easier to represent logical inference on the large-scale text corpora. Our study aims to build a predictive model of news events combining logic patterns. We use the advantages of knowledge representation to predict a subsequent event in generalized logic schemas.

Specifically, we construct evolutionary event ontology knowledge (denoted as EEOK) by semantic analysis tools¹ and represent it into a machine-readable ontology language OWL model for predicting future events. EEOK can be illustrated as a directed cyclic graph with event chains whose nodes stand for generalized events and edges stand for the relations (e.g., temporal, causal) between two events. The ontology knowledge (logic patterns) of evolutionary events can be treated as to generalize a variety of evolutionary event instances, which are specific events and their arguments. To the best of our knowledge, there is no existing evolutionary event ontology knowledge of emergency events in Chinese corpus. Our study proposes solutions to fill the gaps.

It should be noted that both the generalized events knowledge and specific events knowledge are extracted from the same news text. The semantic consistency of the two kinds of knowledge is maintained, thus the error of knowledge alignment is naturally reduced. In this way, EEOK obtains the most appropriate generalization of the occurred event with textual environments into formulated knowledge patterns. These patterns can be injected into the machine independently of natural language by human experts in OWL formalism as a priori knowledge or practical knowledge. The subsequent event can be reasonably inferred based on manually calibrated rules employed as ontology knowledge of evolutionary events with valuable common-sense knowledge. The main contributions of this work are threefold:

(1) We build EEOK that leverages the evolutionary event knowledge represented in standard ontology language OWL with a set of evolutionary patterns.

(2) We use state-of-the-art NLP tools to extract specific evolutionary event instances and their corresponding generalized evolutionary event patterns. Based on merging to constructed EEOK, candidates of events are selected for further prediction. The subsequent event can be integrated into the knowledge base for knowledge updating.

(3) Considering the different event domains, we offer a domain-aware event prediction method since their different patterns existed in different domains. A series of experimental studies have been conducted, and our event domain-aware method has shown superiority performance.

The code of this work is available at https://github.com/RingBDStack/KGEvetPred.

The remainder of this paper is structured as follows: Section 2 describes the related works in TSEP. Section 3 includes the critical components of our model. Section 4 presents the integral system of event prediction. Section 5 provides a problem description. Section 6 describes our model formulation for event prediction modeling. Section 7 provides quantitative and qualitative evaluation, and the last section summarizes our findings and concludes with suggestions for our future work.