Abstract:Most current Event Extraction (EE) methods focus on the high-resource scenario, which requires a large amount of annotated data and can hardly be applied to low-resource domains. To address EE more effectively with limited resources, we propose the Demonstration-enhanced Schema-guided Generation (DemoSG) model, which benefits low-resource EE from two aspects: Firstly, we propose the demonstration-based learning paradigm for EE to fully use the annotated data, which transforms them into demonstrations to illustrate the extraction process and help the model learn effectively. Secondly, we formulate EE as a natural language generation task guided by schema-based prompts, thereby leveraging label semantics and promoting knowledge transfer in low-resource scenarios. We conduct extensive experiments under in-domain and domain adaptation low-resource settings on three datasets, and study the robustness of DemoSG. The results show that DemoSG significantly outperforms current methods in low-resource scenarios.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address the issue of event extraction (EE) in low-resource scenarios. Most current EE methods focus on high-resource scenarios, requiring a large amount of annotated data to achieve satisfactory performance. However, in practical applications, many fields have very limited annotated data due to the high cost and time-consuming nature of annotation. Therefore, these methods are difficult to apply in low-resource domains. To solve this problem, the authors propose a model named DemoSG (Demonstration-enhanced Schema-guided Generation). This model improves EE performance in low-resource scenarios through the following two aspects: 1. **Demonstration-based Learning Paradigm**: Fully utilizes existing annotated data by converting it into demonstrations to help the model better understand the extraction process. 2. **Schema-guided Natural Language Generation Task**: Converts the EE task into a schema-guided natural language generation task, thereby utilizing label semantic information to promote knowledge transfer. ### Specific Methods 1. **Unified Event Record Representation**: Designed a unified template to convert event records into natural language sentences containing event trigger words, arguments, and their corresponding roles. 2. **Event Demonstration Construction**: Uses annotated data not only as traditional supervised learning signals but also converts it into event demonstrations to provide additional information to help the model understand the extraction process. 3. **Schema-based Prompt Construction**: Designed a prompt template that utilizes the semantic information of types and roles in the event schema. 4. **Enhanced Sequence-to-Sequence Generation**: Combines event demonstrations and prompts using an encoder-decoder architecture to generate natural language sentences describing event records. ### Experimental Results The authors conducted extensive experiments under multiple low-resource settings, including in-domain low-resource and domain adaptation scenarios. The experimental results show that DemoSG significantly outperforms existing methods in low-resource scenarios. Specifically: - In the argument extraction task, DemoSG achieved significant improvements under different few-shot and data-limited settings. - In the event detection task, DemoSG also performed excellently, especially in data-limited settings, with the degree of improvement increasing as more data became available. - Different demonstration selection strategies showed different advantages in various tasks, such as the similar retrieval strategy performing better in low-resource argument extraction tasks, while the rich context strategy performed better in low-resource event detection tasks. ### Conclusion By introducing a demonstration-based learning paradigm and a schema-guided natural language generation task, DemoSG effectively improves event extraction performance in low-resource scenarios, providing new insights for solving EE problems in low-resource domains.

DemoSG: Demonstration-enhanced Schema-guided Generation for Low-resource Event Extraction

DEGREE: A Data-Efficient Generation-Based Event Extraction Model

OntoED: Low-resource Event Detection with Ontology Embedding

Schema-based Data Augmentation for Event Extraction

Knowledge Extraction in Low-Resource Scenarios: Survey and Perspective

Boosting Event Extraction with Denoised Structure-to-Text Augmentation

AdaDemo: Data-Efficient Demonstration Expansion for Generalist Robotic Agent

Scale Up Event Extraction Learning via Automatic Training Data Generation

DocEE: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction

One Small and One Large for Document-level Event Argument Extraction

DemoNSF: A Multi-task Demonstration-based Generative Framework for Noisy Slot Filling Task

Good Examples Make A Faster Learner: Simple Demonstration-based Learning for Low-resource NER

Selective Demonstrations for Cross-domain Text-to-SQL

Learning Event Extraction from a Few Guideline Examples

DocEE-zh: A Fine-grained Benchmark for Chinese Document-level Event Extraction

Self-Demos: Eliciting Out-of-Demonstration Generalizability in Large Language Models

KeyEE: Enhancing Low-Resource Generative Event Extraction with Auxiliary Keyword Sub-Prompt

DESED: Dialogue-based Explanation for Sentence-level Event Detection.

STAR: Boosting Low-Resource Event Extraction by Structure-to-Text Data Generation with Large Language Models

Doc2EDAG: An End-to-End Document-level Framework for Chinese Financial Event Extraction

Automatically Labeled Data Generation for Large Scale Event Extraction