Arthasetu - Neural Machine Translation for Classical Jain Philosophical Text with Cross-System Evaluation

Language is often thought of as a bridge between people, but for the world's oldest continuous literary tradition, that bridge has been missing a crucial span. Sanskrit, the sacred language of ancient Indian philosophy, holds a vast library of texts that define the spiritual and intellectual heritage of millions. Yet, in the modern digital age, this language remains largely invisible to the tools that translate the world's languages for us. While computers can now translate casual conversation or news headlines with increasing ease, they struggle profoundly with the dense, highly structured, and compound-heavy sentences found in classical philosophical writings. This gap is particularly acute for Jain philosophy, a major school of thought within the Indian tradition, where the precise meaning of a single word can determine the difference between a spiritual insight and a misunderstanding. The challenge is not just about translating words; it is about teaching a machine to understand a specific, complex way of thinking that has been preserved in a language that few modern speakers read fluently.

To address this, researchers Tamanna Jain and Neeraj Raheja developed a specialized system called Arthasetu, designed to translate classical Jain philosophical texts from Sanskrit into Hindi. Their work focuses on a specific type of artificial intelligence known as a neural machine translation system. In simple terms, this is a computer program that learns to translate by studying thousands of examples of sentences in one language paired with their correct translations in another. The researchers built their system to handle the unique difficulties of classical Sanskrit, which is filled with long, compound words and technical terms that do not appear in everyday speech. They tested their system on a collection of texts including the Tattvartha Sutra, a foundational Jain scripture, and its accompanying commentaries. The goal was to create a tool that could eventually serve as a conversational assistant for students and scholars of Jain philosophy, helping them navigate these ancient texts with greater ease.

The researchers began by gathering and cleaning a massive collection of sentence pairs. They started with a core set of nearly 10,000 sentences drawn directly from Jain philosophical sources. To see if feeding the computer more data would help, they tried two different expansion strategies. In the first, they mixed their philosophical texts with nearly 80,000 sentences from general encyclopedic and contemporary sources, creating a huge dataset of almost 90,000 sentences. In the second, they added only about 5,000 more sentences, but these were carefully chosen to be from other Jain philosophical and devotional texts, keeping the entire collection focused on the same subject matter. The results of this experiment were surprising. The larger, mixed dataset actually made the system worse at translating the specific philosophical texts it was meant to handle. When tested on the Jain-specific sentences, the system trained on the mixed data performed poorly. In contrast, the system trained on the smaller, purely philosophical dataset performed significantly better. This finding suggests that for highly specialized fields, a smaller amount of perfectly relevant data is more valuable than a massive amount of general data.

The team also investigated how the computer should break down the sentences to learn them. A common technique in modern translation systems is to break words into smaller pieces, called subwords, to help the computer handle rare or complex words. However, when the researchers applied this method to their specific dataset, it did not improve the results. In fact, the system performed better when it learned to translate whole words as they appeared. This indicates that for smaller collections of text, the standard tricks used for massive, general-purpose translation systems might not be necessary and could even be a hindrance. The researchers also compared their custom-built system against two major, pre-existing translation engines available to the public. One was a widely used commercial service, and the other was a large, open-source model trained on hundreds of languages. While the commercial service performed the best overall, the researchers' custom system significantly outperformed the large open-source model on the specific task of translating these philosophical texts. This suggests that a system tailored to a specific domain can outperform a massive, general-purpose model that tries to do everything at once.

During their work, the researchers also uncovered a serious flaw in how translation systems are sometimes tested. In an early version of their testing process, the system had a shortcut that allowed it to look up answers directly from the training data during the test, rather than actually translating the text. This mistake caused the system to appear nearly perfect, with a score suggesting it was correct almost every time. Once this shortcut was removed, the scores dropped to a much more realistic level, revealing the true capabilities of the system. This discovery highlights a critical risk in research involving small datasets: if the testing method is not carefully controlled, it can produce misleadingly high results that do not reflect real-world performance. To ensure the system is useful in practice, the researchers propose a hybrid approach for future deployment. This design would first check if a user's question matches a known sentence in their database, returning a verified translation instantly. If the question is new, the system would then use its generative translation skills to construct an answer. This combination of looking up known facts and generating new translations aims to provide both speed and accuracy for users studying these ancient texts.

The work concludes that while the field of machine translation for Sanskrit is still in its early stages, focusing on the specific needs of a sub-domain like Jain philosophy yields better results than trying to apply broad, general solutions. The researchers have demonstrated that a smaller, highly focused dataset is more effective than a larger, mixed one, and that standard methods for breaking down words may not always apply to smaller, specialized projects. They have also provided a clear, reproducible framework for testing these systems, ensuring that future progress is measured accurately. By making their data and methods transparent, they hope to encourage further research into translating the world's classical languages, ensuring that these ancient voices can be heard clearly in the modern world.