CNN-based framework for classifying temporal relations with question encoder.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Temporal relation classification is crucial for understanding text.
• Deep learning models using sentence embeddings are common but lack task-specific information.
• Existing methods struggle with accurately identifying temporal relationships.

Purpose of the Study:

• To propose a novel framework for temporal relation classification.
• To enhance deep learning models by incorporating prior knowledge of time and events.
• To improve the extraction of task-related information from sentence embeddings.

Main Methods:

• Developed a framework incorporating awareness of events and time expressions (time-event entities).
• Utilized a "question encoder" module with various window sizes to focus on context words around entities.
• Extracted task-related information from simple sentence embedding using prior knowledge.

Main Results:

• The proposed framework demonstrated superior performance on the Timebank-Dense corpus.
• Outperformed state-of-the-art temporal relation classifiers, including CNN-, LSTM-, and BERT-based models.
• Effectively leveraged prior information to enhance temporal relation identification.

Conclusions:

• The novel framework successfully addresses limitations in current temporal relation classification methods.
• Incorporating time-event entity awareness significantly boosts classifier performance.
• This approach offers a promising direction for advancing natural language understanding.