Using Clustering Methods for Discovering Event Structures

Cosmin Adrian Bejan, Sanda M. HarabagiuAAAI 2008

Abstract

Traditional natural language processing often treats documents as collections of words, overlooking the structured nature of events that humans use to understand scenarios. This paper presents a method for automatically discovering "event structures"—thematic groupings of related events—from text. By adapting Latent Dirichlet Allocation (LDA) to model event co-occurrences, the authors demonstrate that coherent scenarios such as natural disasters or legal proceedings can be identified in an unsupervised manner. Furthermore, a framework for classifying causal and temporal relations between these events is proposed.

Introduction

In human cognition, information is often organized into scenarios or scripts. For instance, a "crime" scenario involves a predictable sequence: an individual is accused, then arrested, interrogated, and finally brought to trial. Most topic modeling techniques, however, operate at the word level, treating documents as "bags of words." While effective for broad categorization, this approach fails to capture the intricate dependencies—causal, temporal, and hierarchical—between events.

To bridge this gap, this research addresses two primary challenges: identifying which events belong to the same thematic structure (clustering) and determining the nature of their interactions (relation recognition). The goal is to develop a system capable of discovering these structures automatically, reducing the need for labor-intensive manual domain modeling.

Methodology

Event Clustering via LDA

The researchers propose using Latent Dirichlet Allocation (LDA) to discover event structures. While LDA is typically used for word-based topic modeling, here it is applied to events. In this framework:

  • A document is viewed as a mixture of "event structures."
  • Each event structure is defined as a probability distribution over a set of unique events.
  • Events are treated as discrete random variables.

By analyzing how events co-occur across a corpus, the model learns to group them into clusters representing specific scenarios. The granularity of these scenarios can be adjusted by varying the number of clusters ($S$).

Relation Recognition

Beyond clustering, the authors define a taxonomy of six relations to describe event interactions:

  • Subevent: A part-whole relationship.
  • Reason/Purpose/Enablement: Causal or motivational links.
  • Precedence: Sequential or temporal ordering.
  • Related: General connections, such as a reporting event linked to the event being reported.

The authors frame the identification of these relations as a multi-class classification problem, intended to be applied to events within the discovered clusters.

Experimental Setup

The study utilized a custom annotated corpus derived from Google News articles across four domains:

  1. Terrorist attacks (17 articles)
  2. Natural disasters (7 articles)
  3. Deaths of famous personalities (14 articles)
  4. Arrests of criminals (7 articles)

The dataset comprised 3,964 event annotations from 1,262 unique events. Annotation was performed using Callisto for events and a modified Tango tool for relations, adhering to TimeML standards.

Results and Discussion

The qualitative performance of the LDA model (lda-c) was evaluated with the number of clusters set to $S=50$. The model successfully identified clusters that aligned with the target domains in the corpus.

Top events and their probabilities for four discovered event clusters.

As shown in Figure 1, the clusters exhibit high thematic coherence:

  • Cluster 9 (Terrorism): Dominated by events like attack, explosions, and bombings.
  • Cluster 13 (Natural Disasters): Features events such as earthquake, tsunamis, and flooding.
  • Cluster 31 (Death/Funerals): Captures the social and ritual aspects of death, including funeral, eulogized, and grieves.
  • Cluster 47 (Legal/Arrests): Groups events related to law enforcement, such as captured, extraditions, and arrested.

These results suggest that unsupervised statistical models can effectively capture human-centric event scenarios without pre-defined templates.

Limitations and Future Work

While the clustering results are promising, this work is preliminary. The quantitative evaluation of the relation recognition classifier was not fully detailed in this stage. Additionally, the corpus size (45 articles) is small; scaling the approach to larger, more diverse datasets is necessary to test the model's robustness against overlapping or nuanced scenarios. Future efforts will focus on refining the classifiers to automatically detect specific causal and temporal links within the discovered event structures.

Summary updated

BibTeX
@inproceedings{bejan2008using,
  author = {Cosmin Adrian Bejan and Sanda M. Harabagiu},
  title = {{Using Clustering Methods for Discovering Event Structures}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure