AI for the East Frisian Platt
By Thilo Wilts
Artificial intelligence surrounds us almost every day. In many homes Alexa and Siri are listening our commands and kindly wait for their turn to tell a new joke. But these are the visible parts of AI, much more is hidden. For example when we write a text in a messenger like WhatsApp. AI helps us with autocorrect and predicts the next word. Up to today this is only possible in languages like German, English or also Dutch.
We are currently working on an AI that helps us with auto correction and prediction of the East Frisian Platt. For this, we ingest e-books published by our organization member Onno Feldmann, that are hosted on his Website (oostfraeisk.org). Another source is the dictionary on his website. The AI reads these texts and analyses it for frequently connected words. Not only word by word but in a larger context.
Technically spoken a so called Long short-term memory (LSTM), a special kind of neural networks, is trained with the texts. LSTM are well-known for good performance in linguistics.
The problem is the lack of data. Although the previously mentioned e-books are a solid foundation because they are made of natural language, there is not enough text to build a high performing LSTM with good predictions. LSTM usually are trained with vast amounts of data to enable them to understand the language.
Currently we reach an accuracy of ~60% (Oct. 2020). This is not bad for the beginning. But we do not stop here. Our goal is to support the East Frisian Platt and to enable more people to speak the language.
Do not hesitate to contact us: firstname.lastname@example.org