Page 45 - EH75

FACE TO FACE | EASTERN HORIZON 43

understands context, semantics, and generates knowledge of data sources, proficiency in the working
coherent, contextually relevant text responses based on language of the source data and being meditation
Buddhist teachings. In short, NORBU is an AI. practitioners or proficient in Buddhist philosophy.

A chatbot can only provide a response based on So how do we ensure the quality of its knowledge base,
information it receive. Where are your sources of or its ability to wean out “fake quotes”?
information for Norbu?
• All conversations between the bot and users are

To date, more than 3,000 data sources have been used logged and securely stored in the backend system.
to train NORBU. For instance, the following sources have The quality of the bot’s responses is assessed using
been implemented for the Theravada school: markers that allow users to indicate whether the
response was “Helpful” or “Unhelpful.” The “Helpful”
• Entire Pali Tipitaka collection (Sutta, Vinaya, and marks affirm the accuracy of the information
Abhidhamma). These include materials that are provided, while “Unhelpful” responses prompt fine-
already in the public domain, such as Sutta Central tuning efforts to improve future interactions.
(https://suttacentral.net/?lang=en) and Access to • If the bot encounters query to which it cannot
Insight (https://www.accesstoinsight.org/). These provide an immediate answer, it seeks additional
materials are thoroughly debated and rigorously information from users. Feedback received during
checked through online communities such as Dharma these instances is instrumental in refining the bot’s
Wheel (https://www.dhammawheel.com/acc). knowledge base and responses.
• Commentaries from the Pali Tipitaka • Should the bot produces answer that are
• Articles from renowned author-monks and peer- contradictory, due to the way questions are
reviewed academic works. structured, the source guardian will look at them
and will make efforts to “fine-tune” them within the
We are also working closely with CBETA.org from context how the question was framed.
Taiwan and Tibet House based in New Delhi to train • If NORBU is unsure of the answer, it will respond
NORBU on Chinese Mahayana Yogacara / the Agamas with “I’m sorry, I don’t know the answer to that
and Vajrayana source data. question. I am still in the process of being trained.
Can you provide more information, please?”. Strict
A formula of three types of information elements - Right mechanism such as the above are implemented
(authenticity of source), Correct (in terms of versions, to minimize the bot from producing inaccurate
interpretation, and translation), and Verified (by a team responses and to reduce hallucination.
of Dharma experts) - is used to implement the source
data mentioned above to train the bot. NORBU is trained Would Norbu be able to understand the motivations
exclusively with authentic Buddhist source data, in of an individual when questions are asked so that
accordance with the outlined principles. the answers would be more directly applicable and
prescriptive rather than purely descriptive?
How do you ensure that Norbu will not be fed
information that lacks accuracy either in terms of To understand another’s motivations is a complex
translation or doctrine, including fake Buddha quotes? task that requires deep insight and often extended

interaction. While NORBU is able to provide
The administration of the source data is carried out by explanations, offer perspectives, and suggest practices
a full-time team of knowledgeable monastics from all that might be helpful, but the application of these
traditions and lay Buddhists known as “Source Guardians teachings to someone’s specific situation is a personal
of NORBU.” Criteria for selecting individuals as “Source journey. That said, NORBU always aim to provide
Guardians” include expertise in the content, deep answers that are not just descriptive, but also practical