Abstract
We present the first linguistically annotated treebank of Ashokan Prakrit, an early Middle Indo-Aryan dialect continuum attested through Emperor Ashoka Maurya's 3rd century BCE rock and pillar edicts. For annotation, we used the multilingual Universal Dependencies (UD) formalism, following recent UD work on Sanskrit and other Indo-Aryan languages. We touch on some interesting linguistic features that posed issues in annotation: regnal names and other nominal compounds, "proto-ergative" participial constructions, and possible grammaticalizations evidenced by sandhi (phonological assimilation across morpheme boundaries). Eventually, we plan for a complete annotation of all attested Ashokan texts, towards the larger goals of improving UD coverage of different diachronic stages of Indo-Aryan and studying language change in Indo-Aryan using computational methods.
Introduction: Bridging the Indo-Aryan Gap
Computational linguistics has historically focused on high-resource modern languages or prestige ancient languages like Latin and Sanskrit. However, Middle Indo-Aryan (MIA) languages, which serve as the crucial "missing link" between ancient Sanskrit and modern languages like Hindi or Bengali, have remained a significant gap in the field. Ashokan Prakrit, a 3rd-century BCE dialect continuum preserved in Emperor Ashoka Maurya’s famous rock and pillar edicts, represents the earliest attested stage of MIA.
This research addresses the lack of a treebank—a corpus annotated for syntactic structure—for Ashokan Prakrit. By applying the Universal Dependencies (UD) framework, the authors provide a computational foundation for studying the diachronic evolution of South Asian languages as they shifted from synthetic to analytic structures over two millennia.
Methodology and Annotation Framework
The researchers utilized the Universal Dependencies (UD) formalism to ensure cross-linguistic consistency. A notable aspect of their workflow was the use of collaborative annotation via Google Sheets, which facilitated real-time tracking of etymologies and custom features. To handle the phonetic nuances of Prakrit, such as long vowels and aspiration, they adopted a specialized transliteration scheme. Syntactically, the team developed specific dependency subtypes, such as nmod:desc for descriptive nominal modifiers, to capture the unique logic of the language.
The DIPI Corpus and Geographic Variation
The study introduces the Digitizing Imperial Prakrit Inscriptions (DIPI) corpus. Sourced from archaeological records, the corpus currently consists of 12 documents containing 84 sentences and 973 tokens. Despite its limited size, it captures significant regional variation across the Mauryan Empire, including the Girnar (Western), Jaugada (Eastern), and Shahbazgarhi (Northwestern) dialects.
Key Linguistic Insights
The "Regnal Name" Puzzle
A significant challenge involved annotating complex titles like "King Beloved-of-the-Gods Looking-Kindly." While standard UD often uses a "flat" structure for names (treating them as headless lists), the authors argued for a headed noun phrase approach. In this model, "King" serves as the head modified by descriptive titles, more accurately reflecting the internal grammar of Prakrit.
The Transition to Ergativity
One of the most significant findings concerns the -ta participial construction (e.g., "The text was written by me" vs. "I wrote the text"). The authors demonstrate that Ashokan Prakrit exhibits an "ergative-like" transition. In these constructions, the past participle agrees with the object rather than the subject, serving as a clear precursor to the split-ergativity found in modern Hindi.
Grammaticalization and Sandhi
The paper also highlights how sandhi (phonological blending at word boundaries) provides evidence of words evolving into suffixes. For instance, the word for "purpose" (arthāya) began blending with preceding nouns, signaling its transition into a dative case marker—a process that eventually shaped the dative cases in modern Dhivehi and Sinhala.
Results and Evaluation
The project achieved high consistency, with an inter-annotator agreement (Cohen’s $\kappa$) of 0.949 for Part-of-Speech (UPOS) tagging. The Labelled Attachment Score (LAS) was 0.673, reflecting the difficulties inherent in fragmentary ancient texts. Syntactic analysis confirmed that Subject-Object-Verb (SOV) was the dominant unmarked word order, appearing in 50% of transitive sentences, aligning with the patterns of most modern North Indian languages.
Conclusion and Future Work
While the current corpus is a "gold standard" pilot of approximately 1,000 tokens, it sets the stage for a complete annotation of all known Ashokan inscriptions. Future goals include using the treebank to align different versions of the same edicts to study regional variation more deeply and expanding the UD framework to other MIA sources like the Pali Canon and the "Dramatic Prakrits" of ancient Indian theater.
Summary updated