The Key to the City: Using Digital Tools to Understand Tablet Provenience

1 Statistics, corpus linguistics and keywords

Using statistical methods in the comparison of two corpora has been growing in the field of linguistics, specifically corpus linguistics, a methodology of applied linguistics. For example, in this methodology at least one of the two corpora involved in comparisons is a large, standardized bank of English words, such as the British National Corpus (BNC), American National Corpus (ANC) or the Collins Birmingham University International Language Database (COBUILD). These data sets are typically comprised of millions of words collected in the 1980s and 1990s from a broad cross-section of materials. With these large language repositories, researchers have applied various statistical algorithms to better understand language distribution and usage. Linguists compare newspapers, student papers, emails, internet sites, government papers and myriad other contemporary sources to answer questions about significant differences in how people use language. This type of analysis relies upon keywords, that is a lexical item with unusually higher or lower frequency in either the reference or test corpus, and as such is called keyword analysis. These keywords are typically then a starting point for further inquiry into defining a genre or register, identifying communication styles of specific groups/contexts, isolating trends in language usage over time, etc. However, these questions are guided by the contemporary English corpora they draw from. For Assyriology, the application of keyword analysis expands beyond inquiries into sociolinguistics to identifying meaningful similarities or differences between text corpora that can also help address questions about provenience.

The methodology presented here tests the unprovenienced “Diyala” administrative texts ^[3] from the Classical Sargonic period ^[4] against administrative texts from surrounding Diyala sites (Ešnunna, Tutub, Tell Suleimah) as well as the nearby northern site of Kiš from the Classical period. ^[5] Similar Old Akkadian administrative texts from Girsu are used as the control group in order to assess levels of (dis)similarities between sites based on personal names, terminology, language, toponyms, titles, commodities, etc. ^[6]

Keyword analysis alone is insufficient to assign provenience, but, when coupled with well-defined data from other proven techniques, can help elucidate extant analysis. The approach is exploratory rather than focused—it generates hypotheses instead of being guided by them (Gabrielatos 2018: 227). The results presented herein should not be taken as a definitive answer, but as one of several methods of evaluation that can be combined with more traditional methods. As a relatively new tool to Assyriology both the technical and theoretical aspects are outlined below.

2 Keyword analysis

To assist linguists and philologists with statistical analysis of texts there are several software packages available, but AntConc was selected for this study specifically for its multiplatform capabilities, user-friendly interface, gratis price tag and its ability to process transliterated, non-English texts. ^[7] While corpus linguistics has grown beyond English-based corpora, working with non-English texts presents unique challenges in keyword analysis. Fortunately, with the standardized formatting of texts on the Cuneiform Digital Library Initiative (CDLI), corpus creation is relatively easy in Assyriology. In comparing such texts, consistency in spelling, size and genre are crucial (Rayson et al. 2004: 1–2). ^[8]

For Assyriological texts there is often variation in preferred sign readings, which requires some attention before any keyword analysis. Since the word is the unit of analysis, it is mandatory to standardize sign readings that comprise all words. The precise approach adopted is not as important as the consistency of the system. For the Old Akkadian texts, there are two main conflicting approaches to transliteration—that of I. J. Gelb and that of W. von Soden. ^[9] With the Classical Akkadian texts included here, where known, the appropriate value (voiced, voiceless or emphatic) ^[10] was transliterated. This permits the etymological clarity of von Soden’s system but relies on Gelb’s consistency for uncertain readings. ^[11]

For the transliterated data, non-meaningful elements, such as metadata, headers, line numbers, tags, language shift markers and commentary, should be stripped from blocks of texts (downloadable directly from CDLI or created manually). It is particularly important for transliterations of ancient texts to address broken, questionable or emended readings. For example, x-readings and altered “!”-readings for signs were emended to their intended contextual reading (e. g. tum_x → tum₂; zu!(SU) → zu) in order to recover the intended lexeme independent of orthographic variations. ^[12] The reordered sign sequences marked with “:” were normalized and connected with the standard sign connector “-”. Other elements may be erased or altered, depending on the research question and parameters. In this case, since the issue of provenience relies upon levels of (dis)similarity in words within a corpus, all quantities were erased from the text files. And, in order to remove broken passages, AntConc possesses a Stop Wordlist feature that allows the user to define “words” that should not be included in the analysis, such as x-i₃-li₂, …-dingir. ^[13]

An example of this transformation is presented below with the original ATF on the left and the cleaned version on the right ^[14]:

This cleaned data set can now be run through an algorithm to create a ranked raw frequency list of all words within the corpus. However, there may be instances when grammatical variation is not as crucial as other variables; for example, in deciding provenience, it is not particularly relevant whether a verb is in the third person masculine plural or singular. Therefore, words should be lemmatized so that variants can be counted as the same “word.” ^[15] All these measures are taken to try to ensure accurate raw frequency counts for each word in the two corpora, guided by the nature of the specific research question, provenience in this case.

There is no standard statistical method currently for keyword analysis, however each has its own strengths and weaknesses suited to specific purposes (Pojanapunya and Watson Todd 2016: 2). However, all approaches to keyword analysis are based on the frequency of each word in both the reference and test corpus compared against the total number of words in each corpus in order to determine if any difference in relative frequency is due to chance. There are generally two different statistical tests for determining a word’s keyness: significance tests (based on probability statistics) and effect-size tests. Significance test statistics calculate the probability of a frequency difference (i. e. the confidence we can have in the difference being not random) while effect-size statistics focus on the size of the difference between a word’s frequency in the two compared corpora. For the purposes of assigning provenience, this study is more interested in whether the frequency of a word between two corpora can be deemed statistically significant than measuring the size of the frequency difference. Therefore, the probability statistics are preferred.

Probability statistics calculates a p-value that indicates the probability that the difference in a word’s frequency in two different corpora is due to chance. Essentially the smaller the p-value, the more likely the result is not due to chance. The threshold for assigning statistical significance is arbitrary, however the standard threshold for corpus linguistics is p=0.01 (Gabrielatos and Marchi 2012; Gabrielatos 2018: 241).

There are two common probability statistics that produce very similar results in keyword analysis: log-likelihood and chi-square (Pojanapunya and Watson Todd 2016: 13). ^[16] For several reasons, the log-likelihood is preferred. Chi-square becomes unreliable for low frequency words (those occurring fewer than five times) and in small corpora (those with fewer than 50,000 words) (Dunning 1993; Rayson et al. 2004: 3). This has serious implications for results from many Assyriological corpora, including the Old Akkadian texts studied here. Conversely, the log-likelihood is sensitive to corpus size, making results inconsistent across corpora of different sizes (Pojanapunya and Watson Todd 2016: 28). However, this can be controlled for if the level, or threshold for identifying statistical significance, is raised to 0.01 % (=15.13 critical value) (Rayson et al. 2004: 8). ^[17]

Table 1 shows the correlation between the probability of word’s frequency in the two compared corpora and its statistical significance. The p-value corresponds to a critical value that is used as a measure of keyness for each word. A word’s keyness quantifies its uniqueness in either the test corpus (positive values) or reference corpus (negative values). Negative values indicate a high frequency in the reference corpus but a lack of corresponding frequency in the test corpus. Contrariwise, positive values rank words that are more common in the test corpus than the reference corpus. The closer the keyness measurement (critical value) is to zero, the more likely differences in word frequency are merely due to chance; and the fewer words in the keyword list, the more similar the two corpora. Each word (including lemmatized words) generates its own keyness measurement (critical value). And words whose keyness values are above the threshold for statistical significance (15.13) are deemed interesting, important or worthy of further analysis. The interpretation of the results departs from the statistical method into the realm of contextual analysis.

3 The text corpora

The unprovenienced “Diyala” texts serve as the test corpus against which reference corpora from Ešnunna, Tutub, Tell Suleimah, Kiš and Girsu are compared to detect levels of lexical (dis)similarity. The data set is circumscribed by time period and genre in order to control for variation, insofar as is possible with such texts.

The 85 unprovenienced tablets attributed to the “Diyala” corpus here include 51 of the 53 tablets published by Gelb in OAIC, which he generally attributed to the Diyala but without detailed information on their exact provenience. ^[18] This data set is rounded out by a medley of additional texts that have been published in various editions and journals, collected here as a single corpus. ^[19] Some tablets originally assigned a generic “Diyala” provenience have been associated, with some confidence, to Ešnunna through traditional techniques and are, therefore, not included in the “Diyala” corpus here. ^[20]

Using traditional methods, Gelb, P. Steinkeller and J. N. Postgate, and B. Kienast and K. Volk have linked some of these unprovenienced texts with Ešnunna. ^[21] Generally, the presence of personal names from excavated or secure Ešnunna texts, ^[22] specific geographic labels or the use of the deity Tišpak, the city deity of Ešnunna, is invoked as evidence for a tablet’s origin at Ešnunna.

Additionally, Gelb cites the use of specific vocabulary such as šibšum and kušurrā’im (see Table 2). However, despite the similarity in the word choice, the orthography of the same term varies between the excavated Ešnunna texts and those subsequently attributed to the site, leaving the correlation tenuous. ^[23] Therefore, the texts originally assigned to Ešnunna based on this lexical evidence by Gelb are included in the “Diyala” corpus here.

The collection of tablets published in MAD 1 as nos. 270–336 were assigned an Ešnunna provenience by Gelb based on “[b]oth the information given by the dealer from whom the collection was purchased and the internal evidence culled from the tablets,” specifically the co-occurrence of personal names with excavated Ešnunna tablets (Gelb 1952: xi). There are some additional interrelations within this corpus that help assign specific texts to a provenience. The text AuOr 9, 5 mentions i-da-dingir šabra e₂ (“chief administrator of the household”), who is also mentioned with full title in MAD 1, 322, a text confidently associated with Ešnunna. The text AuOr 9, 5 also mentions u-ṣi-um gal-sukkal dingir (“chief sukkal of the deity”), who is present with this same qualifier in JCS 28, 227 (NBC 10,920). A sealing was excavated from Ešnunna with his cylinder seal impression: u-ṣi-um gal-sukkal ^dtišpak, ^[24] demonstrating the reasonable provenience of these two texts to Ešnunna. However, these contextual elements have only limited implications for the remainder of the AuOr 9 texts, which were purchased by P. B. Ubach in Iraq between 1922–1923, possibly in separate lots (Molina 1991: 137). ^[25]

Steinkeller and Postgate have demonstrated the assignment of OrNS 51, p. 355 to Ešnunna and that text’s close relationship to AuOr 9, 4. In a subsequent publication Steinkeller also illustrated the connection between MC 4, 50, JCS 26, 8 and MAD 1, 336 through various land sale transactions of Dabālum (Steinkeller and Postgate 1992: 88–89). ^[26] The internal coherence of these four texts supports an Ešnunna provenience for all four texts given MAD 1, 336’s probable origin from the site.

This medley of unexcavated texts is excluded from the “Diyala” corpus here based on the confluence of evidence suggesting their likely provenience from Ešnunna. This process of defining an unexcavated corpus illustrates the human influence in the statistical outcome and is a part of the process that should continue to be refined and improved. To a point, the human element is unavoidable, however, understanding user-created biases is the first step toward reconciling them.

4 Keyness criteria for estimating provenience

The keyness value may assist in determining how similar the unprovenienced “Diyala” texts are with each of the other five corpora from Ešnunna, Tutub, Tell Suleimah, Kiš and Girsu (see Table 3 for an overview of the corpora).

The unit of analysis is the word, which is defined by lexemes. While certain words have been assigned a lemma, this process is driven by lexical meanings, assigning all observed forms to the same meaning. In the text files, this is denoted by white spaces, which are used to demarcate the boundaries of a given word. This process of defining and delimiting words is subjective and certainly influences the outcome of textual analysis. For this reason, my lemma list, stop wordlist, transliteration files and their generated raw frequency lists are available online so that Assyriologists may access, critique and help improve the user-defined criteria of this methodology. ^[28]

5 “Diyala” texts and Ešnunna

Table 4 below represents those words that demonstrate a 99.99 % statistical significance (15.13 critical value) for uniqueness between the two data sets. The positive values reflect those words appearing in the “Diyala” texts (test corpus) atypically more frequently than the Ešnunna texts (reference corpus). Conversely, the negative keywords, located at the bottom of the table, represent those keywords that occur atypically more frequently in the Ešnunna texts. Based on the frequency of words in the “Diyala” corpus, the software algorithm expected to find similar frequencies for similar words, however there were some significant deviations.

The unprovenienced “Diyala” corpus differentiates itself from the Ešnunna corpus along several lines. First, there is an increased preference for the Akkadian language in the “Diyala” texts (e. g. i-di₃-in, iš-te₄, a-na, ARAD₂-su). Given the preponderance of Akkadian texts in the north compared to southern sites at this time, the “Diyala” texts appear to contain an abnormal number of Akkadian words altogether.

Second, the nature of the economy addressed in each corpus is slightly different. This is an expected deviation given that sites produce goods that are locally viable and/or profitable–the marsh cities producing more fish and reeds, those near irrigated fields produce more grains, etc. But even internally, different archives within the same city may, and often do, focus on different goods. Therefore, the economic differences between the two corpora are not immediately interesting. Despite being “uninteresting,” this type of expected result lends confidence to the statistical analysis.

However, it is important to note that the “Diyala” texts do not utilize the non-Akkadian gur saĝ-ĝal on par with the Ešnunna texts. The Akkadian gur-measure was introduced under the Akkadian kings and was often associated with imperial/royal goods. ^[29] As Foster succinctly concludes about the gur Agade, “only matters for royal accountability were accounted for by the royal standard” (Foster 2016: 49 fn. 86). However, the majority of references of Agade in the Diyala texts is to the toponym, not the capacity measure, suggesting a geographic proximity or other affinity, not necessarily one of metrology. But, coupled with the relative high amount of Akkadian in the “Diyala” texts compared to Ešnunna, a closer relationship with the Akkadian/imperial milieu is posited.

6 “Diyala” texts and Tutub

Similar to the Ešnunna texts, the “Diyala” texts contain more Akkadian than the Tutub corpus, although at a diminished rate (see Table 5 below). The instances of clear Akkadian are more comparable between Tutub and the “Diyala” texts. The specific Akkadian terms that are significantly more prevalent in the “Diyala” texts are linked with differences in economy. The test corpus contains more documents that focus on the exchange rate of grains and silver, while the reference corpus has a more pastoral focus with goats and sheep. Again, this is an expected result, however, the keyness values for these terms are larger than those of the “Diyala”-Ešnunna comparison. This implies that the differences between the “Diyala” and Tutub texts are more pronounced and even less likely to be due to chance.

In the Tutub texts the unclear, yet increased, use of PAP is noted by AntConc as well as the texts’ increased use of patronymics. While the PAP phenomenon is limited and poorly understood, the presence or absence of patronymics is not indicative of any specific site—only more abbreviated texts.

Although the texts from Tutub align in vocabulary with the “Diyala” texts, the differences that are observed are much stronger than those between the “Diyala” and Ešnunna. This might suggest the “Diyala” texts to be an archive at Tutub, dealing with other aspects of the economy or that the “Diyala” texts were written from a more official/imperial perspective at Ešnunna, hence the increase in Akkadian terms, mentions of the city of Agade and limited use of the gur saĝ-ĝal.

7 “Diyala” texts and Tell Suleimah

The statistically significant deviations in the Tell Suleimah corpus are comparatively small in contrast to the sites of Tutub and Ešnunna (see Table 6 below). The prevalence of Akkadian among the Tell Suleimah texts undoubtedly contributes to this level of similarity. Additionally, there are no significant differences in the personal names between the two sites, which could indicate a similar cultural background.

The difference between these two corpora is similar in quality to that of the “Diyala” texts and Ešnunna and Tutub: the difference being one of economy. The different types of grains attributed to each corpus may be due to distinct periods during the agricultural cycle since milled products such as semolina and flours can only be processed after the harvest of barley and emmer wheat.

Again, it is easier to place the “Diyala” texts as part of the Tell Suleimah corpus than of the other two Diyala sites based on the overall similarities in their lexemes.

8 “Diyala” texts and Kiš

When compared with the northern site of Kiš, the similarities between the language, commodities, metrology and personal names is even more striking (see Table 7 below). There are few distinctions between the corpora, and their distribution suggests that the “Diyala” texts could be a subset of the Kiš texts. Based on the higher number of negative results, the words in the “Diyala” texts fit in with the Kiš texts, excepting the use of PAP. However, the Kiš corpora has elements not as prominent in the “Diyala” corpus, suggesting that the Kiš corpus is perhaps broader in content (although not size).

The Akkadian of the “Diyala” texts appears to be most at home among the Kiš corpus, with no statistically significant deviations in Akkadian usage. And similar to Tell Suleimah, there were no significant differences in personal names between those people appearing in the Kiš texts and those in the “Diyala” corpus. Again, this may suggest a shared culture background, if naming practices are in fact similar in nature.

9 “Diyala” texts and Girsu

With Girsu as the control group, it is not surprising that there are enormous deviations between these two corpora in onomastics, commodities, resources, metrology and linguistic affiliation (see Table 8 below). The strength of these results helps situate the unprovenienced texts attributed to the Diyala closer to the corpora of Kiš, Tell Suleimah, Ešnunna and Tutub. There are a relatively high number of words in the “Diyala” corpus that are more unique compared to Girsu. However, given the size of the Girsu corpus, fewer words appear unique for the southern corpus.

Again, there is a relatively high proportion of Akkadian words in the “Diyala” texts compared to Girsu—an expected result. The commodities, implications of the local economies, deviate in similar ways as above—again, an expected difference between sites situated in different ecologies.

Interestingly, there are several “banana” personal names (gi-nu-nu, a-ša-ša, a-li-li, i-bi₂-bi₂, i₃-lu-lu) not as well represented in the southern corpus as in the “Diyala” texts. The linguistic affiliation of this name type remains obscure, but may suggest regional naming preferences by the Sargonic period. ^[30] Certain Semitic names (e. g. Ummi-Eštar, Nabi’um, Bēlī) are expectedly more prevalent in the “Diyala” corpus than the Girsu texts, in accordance with general observations about linguistic distribution during this period. ^[31]

The overall mismatch between these two corpora is expected given the geographic (and thus, ecological, economic, linguistic and cultural) distance of these two sites.

10 Conclusions

There are several possible interpretations of the data, assigning the “Diyala” corpus to a different archive at Tutub, or to a more official archive at Ešnunna, or to an archive written at a different time of year at Tell Suleimah. But the quantified data suggest that despite the possibilities, it is more probable that the “Diyala” texts come from the northern site of Kiš (or one of similar quality). Given both the number of statistically significant words and the size of the keyness values, keyword analysis claims that the lack of dissimilarity between Kiš and the “Diyala” texts are the strongest and least likely to be due to chance. However, this is not the definitive solution to the problem of provenience, but rather a methodology for identifying new areas of inquiry and providing a new vantage on old data. Often, as is also the case here, the results lead to further questions or deeper analysis. The keyword analysis highlights potentially interesting results, but it is still the responsibility of the researcher to assess if and to what degree certain keywords are meaningful. Especially given the unpredictability and irregularity of preservation and discovery, this technique should be paired with the subjective characteristics identified by specialists, such as orthography, paleography, tablet shape and grammatical variation in order to determine probable provenience. Specific differences in language, economy and personal names are each a pathway for deeper exploration of the texts identified by keyword analysis.

In general, however, new methods for determining tablet provenience are particularly relevant with the increase in the number of unexcavated tablets entering collections. Keyword analysis is suited to corpus comparison for larger numbers of cuneiform tablets, complementing the more individual level of analysis of paleography, prosopography, tablet shape, etc. It is important to be able to analyze not just the individual tablet but perhaps an entire lot of tablets to average out quirks, anomalies and random or arbitrary features. As a relatively new methodology, it is my hope that through collaboration this can be refined and improved upon to become a useful tool for Assyriology.