Setswana grammar checker for declarative sentences using LSTM-Recurrent Neural Network

This research is aimed at developing a Setswana grammar checker for Setswana

declarative sentences using Long Short-Term Memory Recurrent neural networks

(LSTM-RNNs). The research was motivated by the fact that Setswana is recognized

as one of the under-resourced languages in the world and the language lacks Natural

language processing (NLP) tools such as grammar checkers; this delays the

language’s technological progress. A Setswana grammar checker is a pre-requisite to

the development of other Human Language (HTL) applications such as machine

translators and parsers that are necessary for the language to exist in the web, hence

contributing to the language’s technological progress or improvement. Various

techniques have been implemented to develop grammar checkers for different

languages. These techniques include the rule-based approach, but the downfall

associated with this technique is that it is language-specific and many rules have to

be developed to satisfy all the grammatical rules available in that specific language;

this may be tedious and time-consuming. Another technique is the syntax-based

approach, and the disadvantage associated with this approach is that it depends on

the availability of a language parser. This research implements the statistical-based

approach to grammar checking. The grammar checker in this research is developed

using Long Short-Term Memory Recurrent neural networks (LSTM-RNNs). The

advantage of this technique lies in the fact that it enables the development of

language-independent grammar checkers and the developer does not need to have

deep knowledge of the underlying grammar of the language they are working with.

The Setswana grammar checker was developed by the use of 1700 Setswana

sentences; 750 incorrect sentences and 750 correct sentences. The training module

had a Validation accuracy of 0.95, a Validation loss of 0.05, and a Training loss of 0.1.

The testing module had a testing accuracy of 0.96. and a testing loss of 0.06. Results

of this study indicate that Long Short-Term Memory Recurrent neural networks (LSTM RNNs) can extract the pattern or word order followed by Setswana sentences and use this information to determine the grammatical correctness of Setswana text as

compared to the rule and syntax-based grammar checking techniques.