When grammar can't be trusted - Valency and semantic categories in North Sámi syntactic analysis and error detection
Linda Wiechetek · NORA - Norwegian Open Research Archives · 2018
In this dissertation, I investigate valencies and syntactically relevant semantic categories in North Sámi. In addition, I develop three machine-readable grammars for the North Sámi grammar checker GoDivvun that have access to valencies and semantics. Like a human, a machine-readable grammar analyzes a sentence by putting together information from different linguistic levels and based on this, selects or discards certain interpretations. Grammatical errors and the extensive homonoymy of well-formed input complicate a reliable sentence analysis based on morphology and syntax alone. I therefore add valency tags to 500 North Sámi verbs and annotate semantic prototype categories to 71% of the noun lexicon. This adds a semantic layer to the sentence analysis that is used to identify governor- argument structures in the process of error detection. I evaluate the detection of a test set of local and global errors resulting in a precision above 98% for local errors and a precision above 77% for global errors. While semantic prototype tagging is the backbone of local error detection, valency annotation is the back- bone of global error detection. My approach shows that a deep syntactic and semantic sentence analysis is beneficial for local error detection and necessary for reliable global error detection.