A Decade of Knowledge Graphs in Natural Language Processing: A Survey

Phillip Schneider, Tim Schopf, Juraj Vladika, Mikhail Galkin, Elena Simperl, Florian Matthes

Introduction

Since Google popularized the term "Knowledge Graph" (KG) in 2012, the intersection of structured knowledge and Natural Language Processing (NLP) has undergone a period of explosive growth. Knowledge Graphs provide the structured, factual "world knowledge" that unstructured text lacks, making them indispensable for tasks ranging from semantic search to complex reasoning. This survey provides a comprehensive mapping of the field, analyzing 507 research papers published between 2012 and 2021 to identify trends, taxonomies, and critical gaps in the current literature.

Methodology: A Systematic Mapping Study

To provide a "big picture" view of a fragmented landscape, the authors conducted a rigorous systematic mapping study. This methodology involved:

  • Search & Screening: Querying six major academic databases, including the ACL Anthology and IEEE Xplore, for papers published in the decade following 2012.
  • Data Extraction: Manually reviewing 746 unique papers to arrive at a final corpus of 507 relevant studies.
  • Multi-faceted Classification: Each paper was categorized by task (using a newly developed taxonomy), research type (e.g., validation research vs. solution proposals), and contribution type (e.g., new techniques vs. new resources).

Taxonomy of Tasks

The survey introduces a unified taxonomy that splits the field into two primary streams:

  1. Knowledge Acquisition: Focused on building and reasoning within KGs. This includes highly studied tasks like Relation Extraction (144 papers) and Entity Extraction (143 papers).
  2. Knowledge Application: Focused on using KGs to enhance NLP tasks. The clear leaders in this category are Question Answering (103 papers) and Semantic Search (91 papers).

Explosive Growth and Geographic Drivers

The integration of KGs into NLP is a contemporary "gold rush." More than 90% of the literature in this field was published between 2017 and 2021. Geographically, research is primarily driven by China (39%) and the United States (23%), followed by India and Germany.

The Dominance of the Health Domain

Among specific application areas, "Health" is the most prominent domain, appearing more than twice as often as scholarly research. KGs are being used extensively in the medical field to navigate complex clinical terminologies and structured patient data. However, the authors note that medical NLP often suffers from poor reproducibility due to data privacy constraints.

The Maturity Gap

A significant finding of the survey is the "maturity gap" between experimental research and industrial application. While tasks like Relation Extraction are highly mature, newer trends like Augmented Language Models—which infuse models like BERT or GPT with structured KG data—remain largely experimental.

Furthermore, the community shows a shortage of "Secondary Research" (2%) and "Evaluation Research" (1%). This indicates that while researchers are prolific at inventing new algorithms, there is a lack of effort toward synthesizing existing knowledge or testing systems in real-world, industrial settings.

The Resource Crisis

The authors observe a lack of new "Resources," such as benchmarks or text corpora. Much of the current research is "overfitting" to a few established KGs, namely ConceptNet, DBpedia, and the Microsoft Academic Graph. This reliance on a small set of datasets may limit the generalizability of new techniques.

Conclusion and Future Directions

The survey concludes that while the first decade of KG-NLP research has been defined by rapid algorithmic innovation, the next decade must focus on maturity. The authors call for:

  • Increased Secondary Research: More surveys and meta-analyses to synthesize the rapid output of the field.
  • New Benchmarks: Moving beyond established KGs to create more diverse and challenging evaluation frameworks.
  • Bridging the Neural-Symbolic Gap: As Augmented Language Models mature, specialized tools will be needed to bridge the gap between neural networks and symbolic knowledge structures.

By addressing these gaps, the community can move from experimental validation toward robust, real-world implementations of knowledge-aware NLP.

Summary updated

BibTeX
@inproceedings{schneider2022decade,
  author = {Phillip Schneider and Tim Schopf and Juraj Vladika and Mikhail Galkin and Elena Simperl and Florian Matthes},
  title = {{A Decade of Knowledge Graphs in Natural Language Processing: A Survey}},
  booktitle = {AACL/IJCNLP},
  year = {2022},
  doi = {10.18653/v1/2022.aacl-main.46},
}
Figure