Mining temporal footprints from Wikipedia

Michele Filannino, Goran Nenadić2014 · DOI

Abstract

Discovery of temporal information is key for organising knowledge and therefore the task of extracting and representing temporal information from texts has received an increasing interest.

Introduction

While we often think of Wikipedia as a collection of structured info-boxes, the vast majority of human knowledge is locked in unstructured prose. For many entities, structured databases like DBpedia may lack a clean "Birth Date" or "Founded Date" field. This research addresses the challenge of temporal footprinting: identifying the continuous period on a timeline that defines a concept's existence.

Examples of temporal footprints for various entities including people, objects, and historical periods.

Knowing that Galileo lived from 1564 to 1642 allows a Question Answering (QA) system to realize that "What computer did Galileo use?" is a nonsensical question. The goal of this work is to approximate these dates for any entity using only its encyclopedic description.

Methodology

The authors propose a three-step pipeline to predict the lower and upper bounds of an entity's existence:

  1. Temporal Expression Extraction (TEE): They compared two methods for finding dates: a simple search for four-digit years (TEE RegEx) and a sophisticated NLP system (TEE HeidelTime) that extracts both explicit and relative dates.
  2. Filtering (Flt): To handle noise (e.g., a 19th-century biography mentioning a 2014 book edition), they used Median Absolute Deviation (MAD) to filter out statistical outliers.
  3. Fitting Normal Distribution (FND): They treated the remaining dates as a probability mass function and fitted a Gaussian distribution to them. They introduced two tuning parameters: $\alpha$ (to control the width of the predicted lifespan) and $\beta$ (to shift the distribution, accounting for the fact that early years of life are rarely mentioned in biographies).

The final bounds are calculated using the formula: $(\mu + \beta) \pm \alpha\sigma$.

Dataset and Evaluation

The researchers evaluated their approach using a massive dataset of 228,824 Wikipedia pages about people born between 1000 AD and 2014. Ground truth birth and death dates were sourced from DBpedia. During preprocessing, they stripped "noise" sections such as "External links" and "Bibliography" to prevent modern dates from skewing the results.

Key Results and Discussion

Performance was measured using Mean Distance Error (MDE), representing the percentage of non-overlap between the predicted interval and the actual lifespan (lower is better).

Strategy Mean Distance Error (MDE) Standard Deviation
TEE RegEx 0.2636 0.3409
TEE RegEx + Flt 0.2596 0.3090
TEE RegEx + Flt + FND 0.3503 0.2430
TEE Heidel + Flt + FND 0.5980 0.2470

The Simplicity Paradox

The most surprising finding was that the simple RegEx approach often outperformed the sophisticated HeidelTime system. This is attributed to the Wikipedia Manual of Style, which discourages vague, relative temporal expressions, making advanced NLP features less relevant for this specific corpus.

The Length Penalty

The researchers observed that the longer a Wikipedia article is, the harder it is to accurately "footprint" the subject. Long articles tend to include more historical context and future references, which act as noise.

The relationship between article length (#words) and Mean Distance Error (MDE) across different extraction strategies.

Conclusion

While the TEE RegEx + Filtering approach emerged as the winner for raw accuracy, methods using Gaussian fitting (FND) showed much lower standard deviation, suggesting they are more robust and less likely to produce wildly incorrect results. Future work could involve optimizing parameters for different entity types, such as historical periods or technological inventions.

Summary updated

Abstract

Discovery of temporal information is key for organising knowledge and therefore the task of extracting and representing temporal information from texts has received an increasing interest.
BibTeX
@misc{filannino2014mining,
  author = {Michele Filannino and Goran Nenadić},
  title = {{Mining temporal footprints from Wikipedia}},
  year = {2014},
  doi = {10.3115/v1/w14-4502},
}
Figure