Construction of Genealogical Knowledge Graphs From Obituaries: Multitask Neural Network Extraction System

Kai He, Lixia Yao, Jiawei Zhang, Yufei Li, Chen LiJournal of Medical Internet Research 2021 · DOI

Abstract

In this work, we discussed the meaning of Genealogical Knowledge Graphs (GKGs) for biomedical research, presented a new version of a corpus with a predefined family relationship map and augmented training data, and proposed a multitask deep neural system to construct and assemble GKGs. The results show our system can extract and demonstrate the potential of enriching Electronic Health Record (EHR) data for more genetic research. We share the source codes and system with the entire scientific community on GitHub without the corpus for privacy protection.

Introduction

While Electronic Health Records (EHRs) are a goldmine for individual health data, they are notoriously poor at capturing family history. Most EHRs only record one or two emergency contacts, and insurance claims data usually only link parents to children under age 26. This lack of a "family x-ray" makes it difficult for researchers to study how diseases like cancer or heart disease move through generations.

Obituaries are a rich, public, but unstructured alternative. They often list multiple generations of relatives, their residences, and their relationships to the deceased. However, manually extracting this data is impossible at scale. The challenge lies in building an AI that can handle the linguistic quirks of obituaries—like maiden names in parentheses or "distributive" last names (e.g., "grandsons Addison and Owen Smith")—without losing accuracy. Researchers have developed a multitask deep learning system that automatically builds complex family trees, or Genealogical Knowledge Graphs (GKGs), by "reading" online obituaries. This system achieves over 93% F1-measure and provides a scalable way to supplement EHRs with multi-generational data for heritability and disease risk research.

Methods

The proposed end-to-end information extraction system is based on a Multitask Learning (MTL) architecture. Unlike traditional "pipeline" models that perform one task after another (which often leads to error propagation), this system shares a common BERT-based encoder across several specialized modules:

  1. Joint NER and RC: Simultaneously identifies names (Named Entity Recognition) and their relationship to the deceased (Relation Classification).
  2. Location Matching: Links specific residences to the correct family members.
  3. Last Name Distributive Parser: A binary classifier that determines if a last name mentioned at the end of a list applies to all preceding names.
  4. Parenthesis Resolver: A 3-class classifier to distinguish if text in brackets refers to a maiden name, a spouse's name, or something else.
  5. Rule-based Inference: A final layer that uses the extracted kinship and names to infer gender, age, and missing birth/death dates.

To handle the "long tail" of rare family relationships (like "granduncle" or "half-sister"), the team used Easy Data Augmentation (EDA) techniques—synonym replacement, random insertion, swapping, and deletion—to ensure the model had enough examples to learn from.

Data and Results

The system was trained on a primary corpus of 1,700 manually annotated online obituaries and extended to a set of 12,407 obituaries collected over a 10-year period (2008–2018). The data was mapped to a predefined family relationship map consisting of 4 entity types and 71 specific relationship types spanning 8 generations.

Distribution of kinship relationships across generations in the constructed Genealogical Knowledge Graphs.

The system achieved a precision of 94.79%, a recall of 91.45%, and an F1-measure of 93.09% on 10-fold cross-validation. The researchers successfully constructed 12,407 GKGs, with the largest single graph containing 30 people across 4 generations. Using a 40% augmentation ratio for rare classes significantly boosted the macroaveraged F1-measure, proving that synthetic "noisy" data can help models generalize better for rare edge cases.

Discussion

One of the most interesting findings was the "Negative Transfer" Phenomenon: including the "Location Matching" module in the joint training actually decreased the performance of the other modules. The authors solved this by co-training highly related tasks (NER, name patterns) while keeping the location task separate.

Obituaries are often more precise than medical records because they explicitly distinguish between biological and non-biological relations (e.g., "son-in-law" vs "son"), which is critical for genetic research. Furthermore, the multitask approach reduced total training time from 240 hours to 150 hours compared to pipeline models.

Limitations and Future Work

  • Selection Bias: Obituaries tend to be published more frequently for affluent families, which could bias genetic studies.
  • Semantic Ambiguity: The system still struggles with ambiguous phrasing, such as "survived by two sons and four grandchildren," where parent-child links between descendants are unclear.
  • Language and Culture: The current system is designed for English-language patterns; expansion would require new linguistic parsers.
  • Privacy: While obituaries are public, aggregating them into searchable knowledge graphs raises privacy concerns. The authors addressed this by masking last names in their shared code.

Summary updated

Abstract

In this work, we discussed the meaning of GKGs for biomedical research, presented a new version of a corpus with a predefined family relationship map and augmented training data, and proposed a multitask deep neural system to construct and assemble GKGs. The results show our system can extract and demonstrate the potential of enriching EHR data for more genetic research. We share the source codes and system with the entire scientific community on GitHub without the corpus for privacy protection.
BibTeX
@article{he2021construction,
  author = {Kai He and Lixia Yao and Jiawei Zhang and Yufei Li and Chen Li},
  title = {{Construction of Genealogical Knowledge Graphs From Obituaries: Multitask Neural Network Extraction System}},
  journal = {Journal of Medical Internet Research},
  year = {2021},
  doi = {10.2196/25670},
}
Figure