DE3TC: Detecting Events with Effective Event Type Information and Context

Neural Processing Letters (2024) 56:89 https://doi.org/10.1007/s11063-024-11570-8 DE3TC: Detecting Events with Effective Event Type Information and Context Boyang Liu1 · Guozheng Rao1,3 · Xin Wang1,3 · Li Zhang2 · Qing Cong1 Accepted: 11 February 2024 / Published online: 6 March 2024 © The Author(s) 2024 Abstract Event Detection (ED) is a crucial information extraction task that aims to identify the event triggers and classify them into predefined event types. However, most existing methods did not perform well when processing events with implicit triggers. And most methods considered ED as a sentence-level task, lacking effective context for event semantics. Moreover, how to maintain good performance under low resource conditions still needs further study. To address these problems, we propose a novel end-to-end ED model called DE3TC, which Detects Events with Effective Event Type Information and Context. We construct an event type-specific Clue to capture the interaction between event type name and trigger words, providing event type information for implicit triggers. For accessing the effective context of event semantics for sentence-level ED, we consider the correlations between types and select similar types’ descriptions as context. With contextualized representation from a contextual encoder, DE3TC learns the event type information for all events including implicit ones. And it performs sentence-level ED efficiently with effective contexts. The empirical results on ACE 2005 and MAVEN datasets show that: (i) DE3TC obtains state-of-the-art performance compared with previous methods. (ii) DE3TC is also excelled under low-resource conditions. B. Liu and L. Zhang have contributed equally to this work B Xin Wang Boyang Liu Guozheng Rao Li Zhang Qing Cong 1 College of Intelligence and Computing, Tianjin University, Tianjin 300350, China 2 School of Economics and Management, Tianjin University of Science and Technology, Tianjin 300457, China 3 Tianjin Key Laboratory of Cognitive Computing and Applications, Tianjin 300350, China 123 89 Page 2 of 20 B. Liu et al. Keywords Event detection · Event type information · Effective context · Low-resource learning 1 Introduction Event Detection (ED) is an essential yet challenging information extraction task in the field of Natural Language Processing. An event is identified by a word or a phrase called event trigger which most represents that event. Given an input text, ED aims to identify the event triggers and classify them into predefined event types. For instance, in the input sentence “Stewart’s 1979 marriage to Alana Hamilton lasted five years and produced two children.”, an ED model needs to recognize the word “marriage” as an event trigger and predict its event type as “Life.Marry”. Early ED methods explored statistical information in the training sets and used patternbased methods [1–3]. Due to the excellent performance of neural network in the field of natural language processing, many ED models used various neural networks to extract contextual semantic features of events [4–6], such as Convolutional Neural Networks (CNN) [4], Recurrent Neural Network (RNN) [5], and Graph Convolutional Network (GCN) [6]. Recently, with the development of the Pre-trained Language Models [7–10] based on the transformer [11] architecture, many powerful ED models have emerged, which better understand the semantics of events in the context and have made significant improvements [12–14]. Some of them regarded ED tasks as question answering (QA)/machine reading comprehension tasks and detected events by finding answers to pre-defined questions [14–16]. For instance, Du et al. [14] formulated ED as a QA task and designed several queries for event triggers. Some generation-based methods manually defined templates or output formats to accomplish event detection [17–19]. They utilized event type information and achieved good performance. Lu et al. [18] designed a sequence-to-structure network and generated different structures for different event types. Hsu et al. [19] proposed a generation-based method and manually designed a prompt containing event type descriptions to guide the process of event detection. According to Liu et al. [20], to better leverage the capabilities of pre-trained language models (LM), various methods reformulate downstream tasks, making them more akin to those solved during the original LM training with the help of a textual prompt. The prompt is derived from the original input using a designed template and serves as the input for the LM. Prompt-based methods reformulate the target task into a generative task, attempting to learn a LM to perform the original task, reducing or obviating the need for large supervised datasets. This process requires additional labor for both task reformulation and template design. There are still three problems as follows that have not been solved: (1) How to detect event with implicit trigger simply but efficiently? In real-life situations, events within the text may often be implicit. In these cases, trigger words do not explicitly convey the semantics of events. For example, as shown in Fig. 1 (Example 1), we can easily tell that the word “marriage” is the trigger for the event type “Life.Marry”. However, in Example 2, it is challenging to determine that “deployed” is the trigger word for the event type “Movement.Transport”. Therefore, we need event type information to identify such trigger words. Most existing methods were unable to handle this situation satisfactorily. The interaction between event types and trigger words provides event type information, which greatly aids in identifying implicit trigger words. Some excellent methods based on QA or machine 123 DE3TC: Detecting Events with Effective Event Type… Page 3 of 20 89 Fig. 1 Two examples of sentence-level event detection from ACE 2005 dataset reading comprehension only provided a keyword as a query for the event type, for example, “[EVENT]”. However, these methods failed to obtain event type information. Some promptbased methods designed different templates for each event type. They may capture the event type information to some extent. But they required a significant amount of manual effort, which is not feasible in real-life applications. (2) How to access effective context of event semantic for sentence-level ED? Context is essential for semantic understanding. The sentence-level event detection task often lacks effective context. It is difficult to identify the trigger words based solely on the semantics of a single sentence. As shown in Fig. 1 (Example 2), the sentence lacks effective contextual information of event semantic. If we provide more event semantics for “Movement.Transport”, the model can identify “deployed” as the trigger word more easily. Furthermore, most existing methods ignored the correlations between similar event types, which provide valuable contextual information of event (...truncated)