Introduction¶
The Ashtadhyayi is a list of around 4000 grammar rules. Together with the Dhatupatha and other secondary texts, it defines a formal grammar that generates valid Sanskrit expressions.
Padmini is an emulator for this grammar that focuses specifically on generating valid words. When feasible, we follow the design and spirit of the Ashtadhyayi and the rest of the tradition. But we balance this adherence to tradition with a pragmatic focus on making a clean and maintainable program.
Why We Created Padmini¶
We created Padmini with several goals in mind:
We want to create a list of Sanskrit verb and nominal forms that is (1) freely available, (2) exhaustive, and (3) high-quality. There are various datasets floating around online, but as far as we can tell, none meets all three criteria.
We want to be able to explain all verb and nominal forms. This ability unlocks powerful new applications. For example, imagine a tool that could explain the grammar details of any word you give it.
We want to create a useful tool for students of Sanskrit grammar. The more comprehensive our tool becomes, the more it can serve as a kind of “reference implementation” that makes sense of the tradition.
More personally, we wanted to dive deeper into the tradition and explore the connection between traditional grammar and formal systems.
As compared to a predecessor system like SanskritVerb, Padmini:
is roughly 10 times faster
fixes several important bugs
has better support for optional derivations
is faster and safer to modify
How We Model Data¶
When grammarians use the Ashtadhyayi, they start with some set of initial conditions and apply rules one after another to produce a final result. This derivation is called a prakriyā. Here is a simple example of a prakriyā for the word “bhavati”:
Each line of the prakriyā is its own step, and it is the result of applying the rule that appears in parentheses. In each step of a prakriyā, there are one or more terms (for example, “bho” or “ti”) that each have their own labels and properties. And these terms consist of some number of sounds that can be modified and changed.
Padmini uses a similar approach. We model a list of sounds as a simple string,
and we associate it with metadata in a Term object. And multiple
Term objects are collected in a Prakriya object, which also
maintains a record of all of the rules that have been applied so far.
In general, Padmini’s user-facing API receives some basic initial conditions and returns all prakriyās that satisfy those conditions.
How We Model Rules¶
Rules are of various types, but at a high level, we can sort them into three types:
interpretive rules, which help us correctly apply other rules. Padmini uses interpretive rules to lay out the program, but these rules are rarely explicit.
modifying rules, which cause some change to the prakriyā. This is the most common rule type that Padmini uses. These rules directly modify the
Prakriya.blocking rules, which block a modifying rule. These rules are logged on the
Prakriya, but they don’t produce any other changes.
A group of rules with a similar purpose is called a prakaraṇa. Based on this concept, we group rules into functions and modules according to their shared intent. This is worth stating explicitly because the Ashtadhyayi itself does not make this choice. Instead, the Ashtadhyayi mainly optimizes for concision and reorders rules if doing so would minimize the total length of the text.
How We Interpret Rules¶
In general, we follow the interpretations of the standard commentaries, especially the Kāśikāvr̥tti and the Siddhāntakauudī. https://ashtadhyayi.com has been an enormous help in this regard.
One small exception is that we have tended to prefer more straightforward interpretations of rules, even if this conflicts with the later tradition. For example, we do not model git in 1.1.5 kṇiti ca. For now, the extra complexity is not necessary.
How We Select Rules for Application¶
Our major departure we make from the tradition is in how we select which rule to use.
The traditional approach is to choose one rule among all of the ones available by using various heuristics. For example, we have this rule from the Paribhāṣenduśekhara:
pūrvaparanityāntaraṅgāpavādānāmuttarottaraṁ balīyaḥ
Earlier, later, obligatory, internally conditioned, and exceptional – these are ordered from lowest to highest priority.
Instead of formalizing the principles of rule selection, we have simply created an ordering for these rules and apply each rule in order if it has scope to apply. This ordering is obviously informed by the tradition, but the underlying mechanics are very different.
Why have we used this approach? Mainly, we’ve done so because it greatly simplifies Padmini’s execution flow and makes it much easier to debug problems and create patches. Once Padmini is stable, we might revisit this choice; but for now, it makes development much faster.