Bidirectional LSTM Tagger for Latvian Grammatical Error Detection

Abstract

This paper reports on the development of a grammar error labeling system for the Latvian language. We choose to label six error types that are crucial for understanding a text as noted in a survey by native Latvian speakers. The error types are the following: an incorrect use of a preposition, an incorrect agreement in a phrase, an incorrect verb form, an incorrect noun form, an incorrect choice of the definite/indefinite ending of an adjective, and a missing comma. For neural network model training, a large amount of error-annotated training data is required. We generate artificial errors in a correct text to cope with the lack of manually annotated data. As a bidirectional Long Short-Term Memory neural network algorithm is considered the best for erroneous word detection by several authors, we chose this architecture. We train several models – models labeling a single type of error and models labeling all six types of errors. The precision for all types of errors reaches 94.61%, the recall – 94.08%.