Abstract:Detecting events and their evolution through time is a crucial task in natural language understanding. Recent neural approaches to event temporal relation extraction typically map events to embeddings in the Euclidean space and train a classifier to detect temporal relations between event pairs. However, embeddings in the Euclidean space cannot capture richer asymmetric relations such as event temporal relations. We thus propose to embed events into hyperbolic spaces, which are intrinsically oriented at modeling hierarchical structures. We introduce two approaches to encode events and their temporal relations in hyperbolic spaces. One approach leverages hyperbolic embeddings to directly infer event relations through simple geometrical operations. In the second one, we devise an end-to-end architecture composed of hyperbolic neural units tailored for the temporal relation extraction task. Thorough experimental assessments on widely used datasets have shown the benefits of revisiting the tasks on a different geometrical space, resulting in state-of-the-art performance on several standard metrics. Finally, the ablation study and several qualitative analyses highlighted the rich event semantics implicitly encoded into hyperbolic spaces.

What problem does this paper attempt to address?

The paper attempts to address the problem of detecting and extracting events and their temporal relationships in natural language understanding. Specifically, existing neural methods typically map events to embeddings in Euclidean space and train classifiers to detect temporal relationships between event pairs. However, embeddings in Euclidean space fail to capture richer asymmetric relationships, such as event temporal relationships. Therefore, this paper proposes a novel approach by embedding events into hyperbolic space to better model hierarchies and temporal relationships. ### Main Contributions: 1. **Hyperbolic Event Embedding Learning**: A lightweight embedding learning method based on the Poincaré ball model is proposed, which directly infers event relationships through simple geometric operations. This method contains only 1.5k parameters, takes about 4 minutes to train, but achieves results comparable to complex models on multiple standard metrics. 2. **End-to-End Hyperbolic Neural Network**: An end-to-end neural network architecture based on hyperbolic geometry is designed for extracting event temporal relationships. This architecture performs excellently on the MATRES and TCR datasets, significantly outperforming existing Euclidean methods. 3. **Experimental Evaluation**: Through thorough experimental evaluation on widely used datasets, the benefits of revisiting the task in different geometric spaces are demonstrated. Ablation studies and qualitative analysis further prove the rich event semantics inherent in hyperbolic space. ### Problems Addressed: - **Asymmetry of Event Temporal Relationships**: Existing Euclidean embedding methods fail to effectively capture the asymmetry of event temporal relationships (e.g., "Event A happens before Event B" but not vice versa). - **Modeling Hierarchies**: Events in text do not simply exist independently but form complex hierarchies. Existing models struggle to effectively model these hierarchies, limiting their performance on real-world datasets. - **Data Scarcity**: The inherent complexity of temporal relationships makes methods that rely solely on annotated data less effective. The introduction of hyperbolic space can better utilize limited annotated data. Through the above methods, the paper aims to improve the accuracy and efficiency of event temporal relationship extraction, thereby achieving better performance in natural language processing tasks.

Extracting Event Temporal Relations via Hyperbolic Geometry

Hyperbolic Temporal Network Embedding

HyperED: A hierarchy‐aware network based on hyperbolic geometry for event detection

Neural Embeddings of Graphs in Hyperbolic Space

Discrete-time Temporal Network Embedding via Implicit Hierarchical Learning in Hyperbolic Space

Extrapolation over temporal knowledge graph via hyperbolic embedding

Survey on Knowledge Graph Embedding Based on Hyperbolic Geometry.

From Semantics to Hierarchy: A Hybrid Euclidean-Tangent-Hyperbolic Space Model for Temporal Knowledge Graph Reasoning

A Geometry-Aware Algorithm to Learn Hierarchical Embeddings in Hyperbolic Space

Representation of Multirelations of Geographic Scenes Based on Hyperbolic Space

Representation of Multi-Relations of Geographic Scenes Based on Hyperbolic Space

Hyperbolic Representation Learning: Revisiting and Advancing

(HTNE)-T-2: Temporal Heterogeneous Information Network Embedding in Hyperbolic Spaces

Event2vec: Heterogeneous Hypergraph Embedding for Event Data

Event causality extraction through external event knowledge learning and polyhedral word embedding

HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces

HyGGE: Hyperbolic Graph Attention Network for Reasoning over Knowledge Graphs

Knowledge Graph Representation Via Hierarchical Hyperbolic Neural Graph Embedding

Embedding Learning with Events in Heterogeneous Information Networks

Hyperbolic Hierarchy-Aware Knowledge Graph Embedding for Link Prediction

Event Temporal Relation Extraction with Attention Mechanism and Graph Neural Network