A Survey of event extraction methods from text for decision support systems

• We identify data-driven, knowledge-driven, and hybrid event extraction approaches.

• A wide variety of decision support applications can benefit from event extraction.

• Pressing research issues to be addressed are scalability and domain dependencies.

• Evaluation with annotated data from standard benchmarks or crowdsourcing is advised.

Abstract

Event extraction, a specialized stream of information extraction rooted back into the 1980s, has greatly gained in popularity due to the advent of big data and the developments in the related fields of text mining and natural language processing. However, up to this date, an overview of this particular field remains elusive. Therefore, we give a summarization of event extraction techniques for textual data, distinguishing between data-driven, knowledge-driven, and hybrid methods, and present a qualitative evaluation of these. Moreover, we discuss common decision support applications of event extraction from text corpora. Last, we elaborate on the evaluation of event extraction systems and identify current research issues.

Introduction

Over the years, Information Extraction (IE) has become increasingly popular as a tool for a vast array of applications [[1], [2], [3], [4], [5]]. At first, the IE field was focused particularly on message understanding in newswires. However, due to the onset of progressively larger digital data collections of various natural language text types such as news messages, articles, and web pages, researchers and practitioners require more advanced techniques, extract more information with greater accuracies and on a real-time basis, and operate on larger scales than ever before. Since the early 2000s, there has been a notable shift from general information extraction from digital collections – extracting basic named entities like persons and organizations – toward more advanced forms of text mining, including Event Extraction (EE) that requires the handling of textual content or data describing complex relations between entities [6]. This development has been fueled by the continuous advances in Text Mining (TM) and Natural Language Processing (NLP), the advent of big data, as well as the availability of (manually) annotated data sets that often serve as a basis for building extraction models.

Event extraction combines knowledge and experience from a number of domains, including computer science, linguistics, data mining, artificial intelligence, and knowledge modeling. It is commonly seen as the TM-aided extraction of complex combinations of relations between actors (entities), performed after executing a series of initial NLP steps. It is a form of IE, aimed at specific users, applications, and platforms, that results in more complex and detailed outputs than regular IE. Event extraction originates in the late 1980s, when the U.S. Defense Advanced Research Projects Agency (DARPA) boosted research into message understanding, aimed at automating the identification of terrorism-related events from newswires, a topic that has remained trending up until today.

With the exponential growth of digital collections and the information extraction requirements in various fields, event extraction research has evolved greatly. Early mentions of modern event extraction can be found in the biomedical literature, where NLP techniques have been traditionally employed for discovering biological entities such as genes and proteins, but where the same techniques are now also widely used for identifying events involving these entities, e.g., gene expressions and protein bindings [7]. Gradually, event extraction has moved to other domains such as politics and finance, where events like elections, CEO changes, or acquisitions, are also comprised of sets of entities (e.g., persons, governments, countries) and their relations (e.g., leadership, competitor, ownership) [8], [9].

The detailed information that is commonly extracted from a heterogeneous set of sources in event extraction implementations, becomes increasingly important for supporting decision making processes. Today, the applications of events in decision support systems are plentiful. For instance, events can be used in mediation information systems [10], for the analysis of firm-specific social media monitoring [11], or even for advanced spatio-temporal reasoning in moving objects [12] and vehicle routing [13]. Other popular applications of events lie in environmental scanning [14], news personalization systems [15], algorithmic trading [16], financial risk analysis [17], e-commerce [18], quality assurance [19], [20], and terrorism detection [21].

Such event-based decision support systems commonly define an event as something that is regarded as happening during a particular interval of time. Events can have multiple occurrences and are generally seen as incidents of substantial importance. In this work, we do not consider organized events such as soccer matches, scientific conferences, and parties, but we focus on unexpected occurrences which need to be acted upon. Such events are universally associated with state changes. However, per domain, their definition, complexity, and interpretation could greatly differ.

Irrespective of their domains, extracted events are associated with changes in the state of the current knowledge, and hence can be employed for decision making, prediction, or monitoring. The applications are numerous, e.g., generating trading signals for stock exchange markets, providing event-driven data integration in decision support systems, creating social media monitoring systems by police departments, and discovering defects in products. Hence, these developments render traders, managers, and companies to be the users that immediately benefit from event extraction.

Despite the envisaged usefulness and wide prospective applicability of event extraction, several hurdles have to be overcome until event extraction is widely adopted as a supportive tool in practice. The main requirements that were trending in the nineties for information extraction [2], are still applicable to event extraction today. For instance, the technologies should deliver sufficiently accurate results. Furthermore, construction and processing costs should be minimized, and systems are preferred to be operable by non-specialists. These challenging requirements have led to many research efforts in the last decade, of which the main ideas are surveyed in this article.

Although IE in general is certainly a heavily researched and well-described area, to our knowledge, there is little overview work focusing on the upcoming field of event extraction. Therefore, in order to aid researchers and practitioners in making well-informed decisions about their event extraction applications, we survey high-performance extraction techniques and their common applications in decision support systems. While a preliminary survey on event extraction from text already exists [22], here we provide a more complete overview on a higher level of abstraction, and also cover the most recent works. Moreover, in our current endeavors, the various approaches to event extraction are evaluated on more (qualitative) dimensions. We discuss common applications of event extraction in decision support systems and additionally focus on the evaluation of event extraction methods. Last, current research issues in event extraction from text are highlighted.