Temporal Information Extraction

Xiao Ling, Daniel S. WeldAAAI 2010 · DOI

Abstract

Traditional information extraction (IE) systems often treat facts as static, neglecting the temporal "fluents" that define the real world. This paper introduces TIE (Temporal Information Extraction), a framework that utilizes Markov Logic Networks and global transitivity reasoning to extract events and precisely bound their temporal start and end points. By moving from interval-based to point-based representations, TIE significantly outperforms existing baselines in both precision and recall.

Introduction

While traditional IE is effective at identifying static relations, such as BornIn(Steve-Jobs, San-Francisco), the majority of human knowledge is inherently temporal. Knowing a fact like "Bill Clinton is President" is only useful if the timeframe is specified. Current research in this area faces three primary hurdles: a bias toward static facts, the use of coarse relations (like "OVERLAP") that lack precision, and a lack of global reasoning to fill in missing temporal information.

Methodology: The TIE Pipeline

The TIE system departs from complex "interval algebras" in favor of a point-based representation. Each event $e$ is defined by two discrete points: a beginning ($\triangleleft e$) and an ending ($e \triangleright$). This simplification allows for the use of linear inequalities and efficient solvers.

The TIE pipeline: from raw text to temporal inference using SRL and Markov Logic Networks.

The pipeline consists of several key stages:

  1. Preprocessing: Utilizing the Stanford parser for syntax and Semantic Role Labeling (SRL) to identify temporal arguments.
  2. Markov Logic Networks (MLNs): MLNs serve as the core reasoning engine, combining first-order logic rules with probabilistic graphical models.
  3. Global Inference: The system enforces transitivity constraints (e.g., if A is before B and B is before C, then A is before C) to tighten event bounds, even when dates are not explicitly stated in the text.

Experimental Evaluation

The system was trained on the TimeBank 1.2 corpus and evaluated using a diverse new dataset extracted from Wikipedia, covering categories such as Warfare, Universities, and Celebrities. To measure performance, the authors introduced Temporal Entropy (TE), a metric that quantifies the "uncertainty" or tightness of a system's temporal bounds.

Precision-Recall curves comparing TIE variants against baselines like PASCA, SRL, and TARSQI.

Key Results

  • Recall Gains: TIE achieved a breakthrough in recall, tripling the performance of baseline systems like PASCA or standard SRL without a significant loss in precision.
  • Bound Tightness: Temporal Entropy analysis showed that TIE’s induced time bounds are the closest to human-level performance among state-of-the-art extractors.
  • Competitive Accuracy: On standard benchmarks like TempEval Task A, TIE achieved an accuracy of 0.695, proving highly competitive with specialized state-of-the-art models.

Discussion and Insights

The success of TIE highlights the power of transitivity (e.g., if $P1 < P2$ and $P2 < P3$, then $P1 < P3$) as a primary driver of recall rather than just a post-processing step. By enforcing consistency across a sentence, the system infers relations that are never explicitly stated. Furthermore, the ablation studies indicate a clear trade-off: SRL features are vital for precision, while transitivity rules are essential for maximizing recall.

Conclusion and Future Work

While TIE represents a significant step forward, it currently operates primarily within single sentences. Future developments will focus on inter-sentence coreference to build document-wide timelines, handling N-ary arguments, and scaling the MLN inference to process web-scale data.

Summary updated

BibTeX
@inproceedings{ling2010temporal,
  author = {Xiao Ling and Daniel S. Weld},
  title = {{Temporal Information Extraction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7512},
}
Figure