Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker (2105.14924v1)

Abstract: Document-level event extraction aims to recognize event information from a whole piece of article. Existing methods are not effective due to two challenges of this task: a) the target event arguments are scattered across sentences; b) the correlation among events in a document is non-trivial to model. In this paper, we propose Heterogeneous Graph-based Interaction Model with a Tracker (GIT) to solve the aforementioned two challenges. For the first challenge, GIT constructs a heterogeneous graph interaction network to capture global interactions among different sentences and entity mentions. For the second, GIT introduces a Tracker module to track the extracted events and hence capture the interdependency among the events. Experiments on a large-scale dataset (Zheng et al., 2019) show GIT outperforms the previous methods by 2.8 F1. Further analysis reveals GIT is effective in extracting multiple correlated events and event arguments that scatter across the document. Our code is available at https://github.com/RunxinXu/GIT.

• The paper introduces a heterogeneous graph-based model with a tracker that effectively captures cross-sentence event dependencies for improved extraction accuracy.
• It employs a novel tracker module that integrates a global memory system to manage interdependencies among multiple events.
• Empirical results show significant F1 score improvements on a large-scale Chinese financial dataset, underscoring the method’s effectiveness in complex scenarios.

Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker

The paper "Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker" presents a detailed exploration into the challenging domain of document-level event extraction (EE). Distinguished from more conventional sentence-level EE, this research tackles the numerous intricacies involved when event data spans multiple sentences within a document. Two main obstacles in document-level EE are identified: the distribution of event arguments across various sentences and the complex interdependencies among events within a document.

The authors introduce the Heterogeneous Graph-based Interaction Model with a Tracker (Git) to address these challenges. The model utilizes a heterogeneous graph to encapsulate interactions between sentences and entity mentions, supported by four specific types of edges. Through these constructions, Git captures global context effectively, which is pivotal when event arguments are dispersed across sentences. Additionally, the Tracker module is introduced to manage and utilize the interdependencies between events by maintaining records actively through a global memory system. This strategic approach allows Git to most accurately extract events of higher complexity.

Empirical results demonstrate Git's superiority over previous methodologies by achieving a notable 2.8 improvement in F1 score on a prominent large-scale public dataset related to Chinese financial documents. This represents an advancement particularly evident in scenarios requiring the extraction of multiple or cross-sentence events, where Git achieves an increased F1 score of 3.7 and 4.9, respectively.

Contributions and Implementation

The paper's key contributions lie in:

1. Heterogeneous Graph Construction: This approach models both entity mentions and sentences, using four types of edges to facilitate interactions—sentence-sentence, sentence-mention, intra-mention, and inter-mention. This formation is crucial for capturing cross-sentence context.
2. Tracker Module Introduction: By incorporating interdependencies among multiple events, Git significantly eases event extraction in multifaceted scenarios through a global memory, aligning the predictive process with other correlated records.
3. Comprehensive Evaluation: On a large dataset comprising 32,040 documents, the Git framework outperforms existing models, affirming its effectiveness especially in documents with complex sentence structures and multiple events.

Implications and Future Directions

This work pushes the boundaries of traditional event extraction beyond single sentences, addressing the need for models adept at understanding events in a document's broader context. It highlights the potential of neural networks augmented with structured and memory components to handle complex linguistic constructs in information extraction tasks.

Future advancements may explore enhancing the scalability and efficiency of Git for even larger, more varied datasets. Another area to explore is the application of Git across different languages and domains, examining its adaptability and generalization capabilities beyond Chinese financial text. Further research could also consider integration with knowledge bases to enrich context further, leveraging broader semantic relations and ontologies to refine event inference. The authors' methodology serves as a promising avenue for improving machine understanding in scenarios where comprehensive context is necessary for accurate information retrieval.