Extracting Chinese events with a joint label space model.

Area of Science:

• Natural Language Processing
• Artificial Intelligence

Background:

• Event extraction involves entity recognition, trigger identification, and argument role classification.
• Current multi-task learning methods treat subtask labels independently, losing valuable information and limiting feature interaction.
• Existing approaches struggle to effectively incorporate label-level interactions.

Purpose of the Study:

• To propose a joint label space framework for enhancing Chinese event extraction.
• To address the limitations of treating event subtask labels as independent one-hot vectors.
• To improve the incorporation of interactive features at the label level.

Main Methods:

• Developed a joint label space framework converting subtask labels into a dense matrix.
• Utilized an incrementally refined attention mechanism for shared label distribution across Chinese characters.
• Incorporated word lexicon into character representation probabilistically to mitigate segmentation errors.

Main Results:

• The proposed model significantly outperforms state-of-the-art methods on Chinese and English event extraction benchmarks.
• Demonstrated the effectiveness of the joint label space in capturing label interdependencies.
• Showcased improved performance by leveraging learned label embeddings and probabilistic lexicon integration.

Conclusions:

• The joint label space framework offers a superior approach to Chinese event extraction compared to existing methods.
• Integrating label information and probabilistic lexicon handling leads to substantial performance gains.
• The model's ability to learn shared label distributions and adjust output layer weights enhances extraction accuracy.