TL;DR
This paper tackles the challenge of deciphering ancient Proto-Elamite (PE) numerals, which are notoriously ambiguous due to multiple overlapping number systems. The authors introduce novel computational methods to automatically extract possible numeral readings and then disambiguate them, significantly improving our ability to understand these ancient accounting records.
What Problem Does This Paper Address?
Ancient scripts, especially those used for administrative purposes, often present a significant challenge for decipherment. The Proto-Elamite (PE) script, used in the early 3rd millennium BCE for accounting, is a prime example. While many of the attested glyphs are digits, the script employs multiple, partially overlapping number systems (decimal, sexagesimal, bisexagesimal, and capacity). This ambiguity means that a single sequence of digits can have several valid interpretations, making it difficult to determine the actual numerical values and, consequently, the content of the accounts. Previous work has largely focused on manual analysis of individual texts, lacking a large-scale quantitative approach to disambiguate these numeral sequences. This paper aims to bridge that gap by developing computational methods to systematically resolve these ambiguities.
Methods
The authors propose a two-pronged approach to disambiguating PE numeral sequences:
- Automated Conversion and Rule-Based Extraction: They first develop an algorithm to automatically extract all possible readings for each PE numeral notation based on the known values of digits within the different number systems. This process leverages structural properties of the original documents.
- Disambiguation Techniques:
- Subset-Sum Analysis: This method exploits the fact that some PE tablets include summary lines at the end. By analyzing these summaries and comparing them to the entries on the obverse, the authors can infer the number system used, especially when the summary or some of the entries are unambiguous.
- Bootstrapping Classifier: They train a classifier using "seed rules" derived from inherently unambiguous numeral notations. A novel "cautious rule selection" approach is introduced to improve the performance of the bootstrap classifier, making it more robust and interpretable. This approach prioritizes rules with stronger evidence, leading to better disambiguation.
Datasets
The primary dataset used is the transliterated Proto-Elamite (PE) corpus hosted by the Cuneiform Digital Library Initiative (CDLI). The authors also contribute a newly constructed test set for evaluating their disambiguation techniques, which was created by manually disambiguating a subset of the ambiguous numerals with the help of domain experts.
Key Insights and Ideas
- Ambiguity is Pervasive: A significant majority of PE numerals are ambiguous, requiring disambiguation to understand their true values.
- Context is Key: The authors demonstrate that structural properties of the documents (like summary lines) and the context in which numerals appear (associated signs) are crucial for disambiguation.
- Mixed Systems are Common: Many accounts unambiguously use two different number systems, often a capacity system (C) alongside an integer system (S, D, or B), suggesting a pattern of recording both goods and their quantities.
- Novel Bootstrapping Approach: The proposed "cautious rule selection" for bootstrapping significantly outperforms a baseline approach, highlighting the importance of carefully selecting rules in this domain.
- Potential for Sign Meaning Inference: The analysis of numeral systems associated with specific signs offers new avenues for inferring the meanings of unknown signs, particularly those related to objects being counted.
- Evidence of Scribe Variability: The presence of "invalid notations" that violate established bundling principles suggests a degree of inconsistency or error in scribal practices, indicating that the writing system may not have been fully standardized.
Key Results
- The automated conversion process successfully extracts possible readings for 7954 out of 8011 intact numerals.
- The subset-sum analysis successfully disambiguates 24 texts.
- The novel bootstrapping approach achieves an F1 score of 0.88 in a 4-way classification task (distinguishing between all four systems), a significant improvement over the baseline's 0.60 F1.
- In a 2-way setting (distinguishing capacity measures from others), their model achieves 0.90 F1, compared to the baseline's 0.69 F1.
- The analysis reveals previously unknown correlations between tablet content and numeral magnitude, offering new insights into the script and its usage.
Limitations and Future Work
- Limited Evaluation Scope: Due to the undeciphered nature of PE, the evaluation is limited to a subset of numerals that could be manually disambiguated by experts. This means the reported performance might represent an upper bound.
- Feature Engineering: The performance of the feature-based classifier is dependent on the chosen features, which, while justified, can sometimes appear arbitrary.
- Broader Decipherment Context: PE numerals are only one part of the larger decipherment problem; results must be interpreted alongside findings from other linguistic and archaeological analyses.
- Future Work: The authors suggest that their methods can be extended to infer the function of signs by examining the variety of number systems they occur beside. They also believe their work is a crucial step towards a deeper understanding of this ancient accounting technology.
Summary updated