Abstract: Network analysis for social disambiguation
By Adam Anderson, University of California, Berkeley
American Oriental Society Meeting 2017
Friday March 17th afternoon - ANE II, Bunker Hill Room
Omni Hotel, California Plaza, 251 South Olive Street, Los Angeles
Already in the early 1960s, the French anthropologist Jean-Claude Gardin had been working on building graphical models to help address specific issues in anthropology (Hymes, Dell ed.1965. The Use of Computers in Anthropology.). After teaming up with the Assyriologist, Paul Garelli, Gardin designed a computational model which they believed would help bridge a major disciplinary divide in the field of Old Assyrian studies concerning the groups of merchants transporting goods between Assur and the Anatolian region in the Bronze Age (c. 1950-1750 B.C.). However, due to the limitations of computational and textual analysis of their day, their results were unable to go beyond a static structural map of these individuals in place and time, and the project was subsequently abandoned. Thanks to the advancements made in recent years, we have been able to return to the same research question of Gardin and Garelli to propose a dynamic solution to their static map.
The research question pursued here, is whether an aggregated approach, combining modern technical analysis (NLP, LDA, and SNA), will allow for an accurate 'distant reading' of ancient documents. The initial text case was 2,000 letters and 3,000 economic, and legal documents from an Anatolian-Assyrian trade colony in central Turkey known as Kanesh (Kültepe). The results of our analysis are used to propose a method for accurately measuring the scale of the society represented in the documentation. In the case of the Old Assyrian society, this method is used for disambiguation, which requires scalable data from the most microscopic (actors on a tablet) to the macroscopic cliques and groups across the network. Scalable data is where Gardin and Garelli's research fell short, and it is here where network analysis, in particular, has proven to be most helpful in managing the intricate details from each text coupled with the large scale of the network (ca. 20,000 nodes). Therefore, our method combines computational tools for textual analysis and a network analysis with graphical interface to inform supervised disambiguation.