[Paper Review] Detecting and Extracting Events from Text Documents

Jugal Kalita|arXiv (Cornell University)|Jan 15, 2016

Topic Modeling176 references3 citations

TL;DR

This paper surveys event detection and extraction from text using machine learning, focusing on identifying events and their participants in unstructured text for applications in summarization, biomedicine, and social media. It proposes EDCoW, a wavelet-based method that detects events by clustering words with bursty, correlated patterns in Twitter data, improving event detection through modularity-driven graph partitioning and cross-correlation analysis.

ABSTRACT

Events of various kinds are mentioned and discussed in text documents, whether they are books, news articles, blogs or microblog feeds. The paper starts by giving an overview of how events are treated in linguistics and philosophy. We follow this discussion by surveying how events and associated information are handled in computationally. In particular, we look at how textual documents can be mined to extract events and ancillary information. These days, it is mostly through the application of various machine learning techniques. We also discuss applications of event detection and extraction systems, particularly in summarization, in the medical domain and in the context of Twitter posts. We end the paper with a discussion of challenges and future directions.

Motivation & Objective

• To provide a comprehensive survey of event detection and extraction techniques in natural language processing.
• To identify challenges in detecting events across diverse domains such as news, biomedical texts, and social media.
• To propose and evaluate EDCoW, a novel method for detecting bursty events in Twitter using signal processing and graph partitioning.
• To demonstrate the utility of event-based representations in summarization and question-answering systems.
• To address limitations in current systems, particularly the inability to detect cross-sentence events.

Proposed method

• EDCoW transforms word frequency signals over time into wavelet representations using a sliding window to convert time-domain data into time-scale domain.
• It applies auto-correlation to filter out trivial, low-significance words before event detection.
• Word similarity is computed via cross-correlation, forming a symmetric sparse correlation matrix to represent signal similarities.
• Events are detected by solving a graph partitioning problem where nodes are words and edges represent cross-correlation, using modularity as the optimization metric.
• The modularity matrix is computed and the largest eigenvalue and corresponding eigenvector are found using the power iteration method for scalability.
• Events are defined as groups of at least two words with high internal correlation and low inter-group correlation, with significance weighted by word count and correlation strength.

Experimental results

Research questions

• RQ1 How can events be effectively detected and extracted from unstructured text using computational methods?
• RQ2 What role do linguistic and philosophical conceptions of events play in designing computational event detection systems?
• RQ3 How can bursty patterns in social media text be modeled to detect meaningful events?
• RQ4 To what extent can event-based representations improve document summarization and question-answering systems?
• RQ5 What are the limitations of current event extraction systems in handling cross-sentence events?

Key findings

• EDCoW successfully detects events in Twitter data by modeling word frequency bursts through wavelet transformation and modularity-based clustering.
• The method improves event detection by filtering low-significance words using auto-correlation and focusing on high-correlation word clusters.
• Event significance is quantified by combining the number of words and their cross-correlation, enabling differentiation between trivial and meaningful events.
• The use of power iteration for eigenvalue computation allows EDCoW to scale efficiently with increasing word counts in large text streams.
• Event detection via graph partitioning using modularity outperforms random edge assignment, as shown by the modularity metric's ability to measure internal cohesion and external separation.
• The system demonstrates that event-based representations enhance summarization and question-answering, as shown by improved performance over baseline bag-of-words models.