Tulu-Based Algorithmic Language Model

Tulu is an Indian language with many speakers but few usable data sources. The 3-source model uses verified data along with data from users to build an understanding of the Tulu language. It implements a modified Hidden Markov model to weigh Part Of Speech (POS) sequences to decide which sequences are yet to be understood or which sentences are more likely. Past attempts to build models for low-resource languages have mainly attempted to use similar languages which are high-resource, or else used rule-based translation. These have a moderate amount of success with regards to accuracy when compared against high-resource models. Efforts to combine the two methods have also been made. A combination of a limited database, user inputs, and community feedback is employed in the 3-source model to create a language model. This model can be replicable for other languages. The project has three intended outcomes: one, to develop a dataset for Tulu words and sentences in a usable format; two, to be able to generate and translate sentences which are accurate in terms of syntax and semantics; and three, to be able to infer results from user feedback and improve its understanding of Tulu. Possible applications of the system include preserving languages having small communities, document translation, language learning applications, and simplifying error correction in translation. The system provides a weighted average accuracy of 24.42 % using Word Error Rate (WER) as the evaluation metric.