Investigating antiquities trafficking with generative pre-trained transformer (GPT)-3 enabled knowledge graphs: A case study
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
<ns3:p> <ns3:bold>Background:</ns3:bold> There is a wide variety of potential sources from which insight into the antiquities trade could be culled, from newspaper articles to auction catalogues, to court dockets, to personal archives, if it could all be systematically examined. We explore the use of a large language model, GPT-3, to semi-automate the creation of a knowledge graph of a body of scholarship concerning the antiquities trade. </ns3:p> <ns3:p> <ns3:bold>Methods:</ns3:bold> We give GPT-3 a prompt guiding it to identify knowledge statements around the trade. Given GPT-3’s understanding of the statistical properties of language, our prompt teaches GPT-3 to append text to each article we feed it where the appended text summarizes the knowledge in the article. The summary is in the form of a list of subject, predicate, and object relationships, representing a knowledge graph. Previously we created such lists by manually annotating the source articles. We compare the result of this automatic process with a knowledge graph created from the same sources via hand. When such knowledge graphs are projected into a multi-dimensional embedding model using a neural network (via the Ampligraph open-source Python library), the relative positioning of entities implies the probability of a connection; the direction of the positioning implies the <ns3:italic>kind</ns3:italic> of connection. Thus, we can interrogate the embedding model to discover new probable relationships. The results can generate new insight about the antiquity trade, suggesting possible avenues of research. </ns3:p> <ns3:p> <ns3:bold>Results:</ns3:bold> We find that our semi-automatic approach to generating the knowledge graph in the first place produces comparable results to our hand-made version, but at an enormous savings of time and a possible expansion of the amount of materials we can consider. </ns3:p> <ns3:p> <ns3:bold>Conclusions:</ns3:bold> These results have implications for working with other kinds of archaeological knowledge in grey literature, reports, articles, and other venues via computational means. </ns3:p>
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.006 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.004 |
| Science and technology studies | 0.002 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it