CodEOE: A Benchmark for Jointly Extracting Cross-Document Events and Opinions from Social Media

ACL ARR 2025 May Submission2051 Authors

Abstract: Event extraction (EE) and opinion or sentiment analysis have been extensively studied within recent decades, but their joint research remains an under-explored area. To bridge the gap in event-level opinion and sentiment analysis, we introduce the Cross-Document Event-Opinion Extraction (CodEOE) task, which aims to capture complex event-opinion interactions and jointly extract event triggers, event arguments, opinions and sentiment polarities towards events from multiple documents. The CodEOE task requires a model extracting trigger-argument pairs and trigger-opinion-sentiment triplets by understanding cross-document contexts. We manually construct a high-quality bilingual CodEOE dataset in both Chinese and English with 6,000+ trigger-argument pairs and 4,000+ trigger-opinion-sentiment triplets. We develop an end-to-end model based on the grid-tagging method to benchmark the task, which can effectively perform cross-document context understanding and achieve pair and triplet prediction. The results of our model surpass those of two strong baselines and are comparable to large language models. We hope that this new benchmark will advance research on event-level opinion and sentiment analysis. Our data and code are available \href{https://anonymous.4open.science/r/CodEOE-08BD}{here for peer review}.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: Event Extraction, corpus creation; benchmarking

Contribution Types: Data resources

Languages Studied: English, Chinese

Submission Number: 2051