Abstract
This study presents a pattern-based bootstrapping approach to automate the extraction of historical profiles—including titles, hometowns, and social relations—from Classical Chinese historiography. By leveraging a small set of seed patterns and a reliability-ranking algorithm, the system reconstructs the lives of Han Dynasty figures, overcoming the challenges of low-resource language and linguistic complexities like Zero Pronouns. The method demonstrates significant improvements in precision and recall without the massive annotated datasets typically required by deep learning.
Introduction
For historians, mapping the social networks and career paths of figures in early Chinese empires (c. 841 BC – 1911 AD) is a monumental manual task. While modern Natural Language Processing (NLP) could theoretically automate this, Classical Chinese presents a "low-resource" challenge. Standard deep learning models require millions of labeled examples, which do not exist for 2,000-year-old texts.
Furthermore, the language is linguistically complex: it lacks white spaces, uses a wide array of naming conventions (first names, courtesy names, and titles), and frequently employs "Zero Pronouns" (ZP)—where the subject of a sentence is omitted because it is implied by context. This paper bridges the gap between digital humanities and NLP by creating a system that learns to extract information with minimal human intervention.
Methodology
The authors propose a pattern-based bootstrapping method. Instead of training a massive neural network, the system starts with just one or two "seed patterns" (e.g., [Relegated to $Title]). The process follows a three-step iterative loop:
- Candidate Generation: The system analyzes characters surrounding known entities to find new potential patterns. In Classical Chinese, target values often appear at the end of a clause, so the focus is on windows of characters immediately preceding the value.
- Ranking: To filter unreliable patterns, the authors developed an estimation score ($r(p)$) based on textual similarity (using Hamming distance) and variety. A pattern is considered more reliable if it extracts a diverse set of true values rather than repeating the same instance.
- Expansion: The top-ranked patterns are used to find new entities, which are then fed back into the next iteration to discover even more patterns.
Datasets and Evaluation
The study focuses on two seminal works of Chinese history: the Records of the Grand Historian (Shiji) by Sima Qian and the Book of Han (Hanshu) by Ban Gu. To evaluate the system, domain experts (history professors) curated a ground-truth dataset including 50 handcrafted extraction patterns and 15 complete "gold standard" person profiles containing 158 specific attribute values, such as father-son and master-disciple relationships.
Key Insights
The "First Clause" Assumption
To address the Zero Pronoun (ZP) problem, the authors adopted a heuristic from historical research: if a person is identified in the first clause of a paragraph, all subsequent omitted subjects in that paragraph are assumed to refer to that same individual.
The Bootstrapping Barrier
A fundamental limit to bootstrapping was discovered. The method works effectively for "one-to-many" attributes like Titles, as many people can hold the same title, allowing the algorithm to "hop" between different patterns. However, it struggles with "one-to-one" relations like Father-Son pairs because each pair is unique in the text, providing no common instances to link different patterns together.
Reliability vs. Frequency
The research highlights that frequency does not equate to accuracy. Some patterns appear often but are "noisy," while others are rare but highly accurate. The ranking formula specifically penalizes patterns dominated by a single, repetitive value to ensure broader discovery.
Results and Conclusion
Using patterns designed by history professors, the system achieved an F1 score of 0.851 on the ground-truth profiles. For the automated "Person-Title" extraction task, the bootstrapping method significantly improved over time, peaking at iteration #7 with an Area Under the Curve (AUC) of 0.207 (up from an initial 0.107). Overall, the system increased Average Precision by 42.55% after just three iterations.
While the method currently faces an "instance-level" bottleneck for unique social relations, it provides a powerful framework for digital historiography. Future work will involve integrating more sophisticated linguistic rules and variable window sizes to further refine the extraction of complex historical profiles.
Summary updated