Do two clay fragments bearing similar signs belong to one object? They might preserve separate copies of the same text, or belong to one tablet without their broken surfaces touching. While ranking candidates connected with medical texts, I found that the notes attached to the fragments were essential to distinguishing these possibilities.
The eBL/Nineveh work left a list of 75 candidate pairs. An inspection of the existing notes for the first 20 pairs exposed descriptions of known relationships and non-contiguous joins. A list intended to help find a new join needed to carry what researchers already knew about each fragment alongside its score.
Different scores measured different kinds of similarity
The candidate table kept text similarity separate from differences in object dimensions. Its text input was transliteration: cuneiform writing represented in a consistent notation. The fields compared the overall sign content, the overlap between sign sets, and the overlap in the most similar line.
The whole-text comparison used TF-IDF similarity, which adjusts the weight given to commonly occurring signs. Sign-set and line comparisons used Jaccard similarity, the proportion of shared elements in the combined set. A strong match in one line could coexist with weak similarity across the whole transcription.
One stored pair illustrates the difference. Its maximum line Jaccard value was 1.0, while its whole-text TF-IDF similarity was approximately 0.097. The model score on the same row was approximately 0.866. These measure line-level sign overlap, overall text similarity and the model’s ranking score respectively. There is no basis for reading the last value as an 86.6% probability of a physical join.
The table also included absolute differences in width, length and thickness. Some thickness differences were NaN, indicating that the calculation had no value. Treating missing thickness as a difference of zero would make an unmeasured comparison look favourable. The evidence here supports ranking by text and dimensions; it does not demonstrate a validated image model that fits fracture surfaces together.
What the notes changed
I placed the notes and publication information for both fragments beside the candidate scores. Some notes described previously identified indirect joins; others reported that no physical contact had been found. Another mentioned a comparison fragment while explicitly excluding a join. These were records made by the researchers who documented the objects, not new expert judgments produced by this project.
A statement in one fragment’s notes could not automatically become a verdict on the current candidate pair. It might refer to a different object, so the relationship itself needed reading. Even so, checking known joins and textual correspondences was necessary before presenting any candidate as something newly found.
One highly ranked row marked one fragment as medical while the other fragment’s notes described a hymn. That contrast alone could not rule out a relationship, but matching signs were clearly insufficient to settle it. The pair still needed to be considered alongside its images, transliterations and existing readings.
What the list can support
A package was also prepared for image review, and the saved execution record lists comparison-image files extracted from it. Assembling those files did not establish that the broken surfaces fit. The work described here reached candidate ranking and inspection of existing notes; it did not establish a new physical join.
Reusing this list would first mean checking how known joins were divided between training and evaluation. The next review would record candidate-level disagreements in text, dimensions and fracture surfaces. The 75-pair list provides places to begin that examination.