Clustering Indus Texts using K-means

Abstract

One of the most important undeciphered scripts of the ancient world is the Indus script. Earlier studies had focused on the correlations between signs in the Indus texts using various statistical and computational techniques such as N-grams or Markov chains. In the present study, K-means clustering, an unsupervised machine learning technique is used to identify clusters of similar texts without making any assumptions about its content. The technique is effective in extracting significant clusters and patterns in the script. Nine clusters are extracted from this study. The texts in each cluster share a common set of structural elements and are more similar to each other than the texts in other clusters. The clusters, as extracted from the study, reveal inherent patterns due to adjacent and non-adjacent dependencies between signs in the Indus texts. These clusters have definitive patterns in the usage of the signs but are only weakly associated to any archaeological site or medium of writing. The characteristic signature features of each cluster are identified in the study.

Statistical Analysis of the Tables in Mahadevan’s Concordance of the Indus Valley Script

“…However, there are other systems with several hundred signs in their inventories, such as the Old Cretan "hieroglyphic seals", the Naxi Dongba symbols from Yunnan Province in China, and the Rongorongo from Easter Island., but this fact does not automatically translate them into mature phonetic writing systems. Most scholars think the Indus script is writing (Parpola, 1994;Vidale, 2007;Rao et al, 2009;Yadav et al, 2017) and the balance of opinion favours a Dravidian rather than a Sanskritrelated underlying language (Robinson, 2016). Mahadevan himself believes that the Indus script represents an early form of Dravidian (Robinson, 2009: 276).…”

“…Most scholars tend to think the signs are linguistic, but Rao (reported by Locklear, 2017) believes that until we find longer sequences of signs or a multilingual text, statistical strategies are our best hope for understanding more about the Indus signs. Yadav et al (2017) found that inscribed objects found in different regions have noticeably different sign sequences, such as when comparing finds from Iraq with those from India/Pakistan. They also found some association between sign sequences and the medium of writing.…”

Structure of Indus Script

“…machine learning, pattern recognition, information retrieval and image analysis. Using a clustering technique (K-means) we explored the noncontiguous associations between the signs in the Indus texts and clustered the Indus texts such that the texts in each of these clusters were more similar amongst themselves than to texts belonging to other clusters (Yadav et al, 2017).…”

“…The technique of clustering is used in several fields including Table 4. Frequent sign quadruplets (contiguous/non-contiguous) in clusters C1 to C9 (Yadav et al 2017). variation in the Indus writing on objects found at various sites and also across different types of objects in Yadav 2013.…”