TCG-Event: Effective Task Conditioning for Generation-based Event Extraction

TCG-Event: Effective Task Conditioning for Generation-based Event Extraction Fatemeh Shiri, Tongtong Wu, Yuan-Fang Li, Gholamreza Haffari Department of DS&AI, Faculty of Information Technology, Monash University, Melbourne, Australia fatemeh.shiri, tongtong.wu, YuanFang.Li, Gholamreza.Haffari@monash.edu Abstract Event extraction is an important but challenging task. Many existing techniques decompose it into subtasks of event and argument de- tection/classification, which are themselves complex structured prediction problems. Generation-based extraction techniques lessen the complexity of the problem formulation and are able to leverage the reasoning capabilities of large pre-trained language models. However, the large diversity of available event types makes it hard for generative models to effec- tively select the correct corresponding templates to predict the structured sequence. In this paper, we propose a task-conditioned generation- based event extraction model, TCG-Event, that addresses these challenges. A key contribution of TCG-Event is a novel task conditioning technique that injects event name information as prefixes into each layer of an encoder- decoder-based language model, thus enabling effective supervised learning. Our experiments on two benchmark datasets demonstrate the strong performance of our TCG-Event model. 1 Introduction Event extraction (Li et al., 2021a) aims at extracting structured event records from unstructured text. For example, as shown in Figure 1, event extraction aims to map the sentence “Two homemade pressure-cooker bombs are detonated remotely, killing three and injuring some 260 others” to four predefined event types, e.g., <event type: explosion, trigger word: detonated, role:bomber: Tsarnaevs, ..., role:bomb: homemade pressure-cooker bombs, role:place: Boston Marathon>, as well as other events that are triggered by words killing, injuring and lost limbs. Event extraction is challenging because of the diversity of natural language expressions and the complexity of event structures. These challenges become even more severe in sentences in which the Figure 1: An illustration of the event extraction as a structured generation, with introducing the event type as the task-conditioning to prefix, the decoder can selectively generate sequentialized event representation, which is mentioned in the input text. text generally contains more events. Currently, most event extraction methods employ a decomposition- based approach (Xu et al., 2021), i.e., decomposing the structured prediction problem of a complex event into classification over substructures, such as trigger detection, entity recognition, and argument classification. Many of these methods tackle the subproblems separately, which requires additional annotations for each stage (Paolini et al., 2021). Furthermore, designing an optimal composition ar- chitecture of different subtasks is very challenging. Natural language generation techniques have been successfully applied to several NLP tasks (Raf- fel et al., 2020; Li et al., 2021b; Hsu et al., 2022). They have inspired the use of controlled event generation to tackle event extraction. These ap- proaches use manually designed templates to wrap input sentences and train a model for cloze-style filling. The study by Lu et al. (2021) proposes to generate linearised event records via a pretrained encoder-decoder architecture combined with a constrained decoding mechanism that alleviates the complexity associated with template combination when extracting multiple events. The advantage of the approach of extraction-as-generation is the removal of the need …

Content truncated to 500 words