The Big Breakthrough in Colombo with ChatGPT
How would ChatGPT work with local languages in Sri Lanka? ChatGPT’s ability to run with local languages could make a huge difference to the local communication industry. I warmed up to ChatGPT with a prompt in the Sinhala language (first).
The prompt translated to: “Get data on the spread of cancer in Sri Lanka. Write a Facebook Ad using this data targeting the youth.”
The Sinhala language results were returned within seconds, but unfortunately, the output was dismal. It seemed that ChatGPT struggled to find data created in Sinhala characters, most likely due to a very limited data set.
Still, it is important to see that, despite its limited access in Sinhala, ChatGPT could read Sinhala and even process responses in Sinhala though it was incoherent.
A simplified question in Sinhala, emphasizing data / research by ChatGPT, was fed thereafter.
“Name the nine Provinces of Sri Lanka.”
The results came in within 10 seconds in the Sinhala language, but once again, they were dismal. The output included places such as Colombo, Galle, Kalundas (?), Kannapathindha (?), Matale, Matara, Nuwara (Kandy), Dhakunugama, and Battaramulla).
Still, the results were produced in a grammatically acceptable language style. Again, despite its limited training in Sinhala, ChatGPT could read Sinhala and even process, it to some extent.
Then a request was made in the Tamil language. “What are the leading Tamil language newspapers
published in Sri Lanka?”
The answer came in within 5 seconds in the Tamil language. While ChatGPT is capable of understanding Tamil inputs and responding appropriately, the accuracy of the information it produces needs improvement. During testing, ChatGPT listed leading newspapers published in Tamil Nadu, India, rather than Sri Lanka, indicating that its Tamil language data set requires further refinement.
Another request in Tamil: “Write a short advertisement in Tamil language explaining how meditation and yoga promote good health of individuals (the advertisement request was generic and was not specified such as a “Facebook Ad” etc).
The response received, in the Tamil language this time, sounded more to the point.
Translated to English (via Google Translator), it read: “Meditation and Yoga are considered two religious achievements. While meditation is the practice of attaining contentment in the mind, yoga is the practice of not diminishing the states of the body. This Facebook Ad has a lot of clarity to make your mind up, meet and get healthy. By systematizing meditation and yoga practices, you can reach a higher state of contentment. It will help you to get healthy.”
Interestingly, ChatGPT was not asked to write an ad for Facebook. It was asked only to “write an ad” though it responded “This Facebook ad has a lot of clarity”, which showed there is still some confusion with ChatGPT’s Tamil language LLM processing, despite it speedily catching up on Tamil LLM training.
Conclusion? A few attempts are not sufficient to draw conclusions on ChatGPT’s capacities in Tamil and Sinhala.
The Tamil language processing appeared to show an improved Tamil LLM at work, while the data set needed more work.
As for the Sinhala language processing (whose vocabulary corpus is smaller) it needs “more work” in its LLM aspect and even the data set.
Still, it is encouraging to see that ChatGPT can identify and process Sinhala inputs as Sinhala language inputs.
This, is a good start.
