TL;DR
This paper introduces ExcavatorCovid, a system that automatically extracts COVID-19 related events and their temporal and causal relationships from vast amounts of text data. It then visualizes these relationships in a Temporal and Causal Analysis Graph (TCAG), enabling policymakers and analysts to better understand the pandemic's complex dynamics and potential downstream effects.
What Problem Does This Paper Address?
The COVID-19 pandemic generated an overwhelming volume of information from news articles and scientific publications. This deluge of data made it incredibly difficult for policymakers to track events, understand their causes and impacts, and make timely, informed decisions. Existing methods for analyzing such large-scale textual data were insufficient to provide a comprehensive, interconnected view of the pandemic's unfolding events. ExcavatorCovid aims to bridge this gap by providing a structured, analyzable representation of this information overload.
Methods
ExcavatorCovid employs a multi-stage natural language processing pipeline:
- Event Taxonomy Development: A hybrid approach combining manual curation and automated clustering of verb and noun phrases from a large news corpus was used to build a comprehensive taxonomy of 76 COVID-19 related event types.
- Event Extraction: A BERT-based sequence tagging model is used for two main tasks:
- Event Classification: Identifying event trigger spans and classifying them into the defined taxonomy.
- Event Argument Extraction: Identifying location and time arguments associated with each event mention. This structured information is crucial for targeted analyses.
- Temporal and Causal Relation Extraction: A combination of pattern-based and neural network approaches (using BERT) is employed to extract three types of relations: Causes, Mitigates, and Before. The system merges subtypes for a more user-friendly output.
- TCAG Construction: All extracted events and relations are aggregated into a Temporal and Causal Analysis Graph (TCAG). This graph is visualized through an interactive web-based interface.
- Event Popularity Measurement: A method to measure event popularity over time is implemented, allowing for trend and correlation analysis of events. This involves calculating a normalized popularity score for each event type per month.
Datasets
- Training Data:
- A portion of the Aylien Coronavirus News Dataset was used for event taxonomy development and training the event classification model.
- The ACE 2005 dataset was used for training the event argument extraction model (focusing on location and time arguments).
- A blended dataset consisting of the Entities, Events, Simple and Complex Cause Assertion Annotation datasets from LDC, along with 1.5K temporal relation instances generated from Gigaword articles, was used for training the relation extraction model.
- Corpora for TCAG Production:
- 1.2 million articles from the Aylien Coronavirus News Dataset (filtered for January-May 2020).
- The COVID-19 Open Research Dataset (CORD-19), containing over 300,000 scholarly articles from sources like PubMed, WHO, bioRxiv, and medRxiv.
Key Insights and Ideas
- The Power of Structured Information: The core idea is that by transforming unstructured text into a structured graph of events and their relationships, complex phenomena like a pandemic can become more understandable and actionable.
- Hybrid Approach for Taxonomy: Combining manual expertise with automated clustering is an effective way to build domain-specific ontologies that are both comprehensive and relevant.
- Temporal and Causal Links are Crucial: Simply identifying events isn't enough; understanding when they happened and why (or how they influenced other events) is key to insightful analysis.
- Visualizing Complexity: A TCAG provides a powerful, intuitive way to grasp the interconnectedness of events, revealing both expected and unexpected causal pathways.
- Beyond News: Integrating scholarly articles alongside news provides a richer understanding, capturing both real-world events and deeper analytical insights.
Key Results
- Excavator extracted 6.2 million event mentions across 59 types from the combined corpora.
- The event extraction accuracy was found to be 83% after manual review.
- The system extracted 226,176 causal and temporal relations with a precision of 0.78 for "Causes" and 0.87 for "Mitigates."
- The TCAG visualization allows users to explore event popularity trends and correlations, such as the strong correlation between "Lockdown" and "EconomicCrisis."
- Targeted analysis at the geolocation level (e.g., by US state) was demonstrated, showing how policies like lockdowns varied in impact across different regions.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the event taxonomy is a "partial illustration" and that the system's accuracy is based on manual review of a sampled subset.
- The current system focuses on location and time arguments, with other roles being ignored.
- Future Work:
- Integrate Modal Dependency Parsing for event factuality assessment.
- Develop cross-lingual transfer learning capabilities to extend Excavator's applicability to more languages.
- The authors suggest that the system's utility extends beyond the COVID-19 pandemic, empowering analysts to tackle complex problems in the future.
Summary updated