Negation is a pervasive phenomenon in natural language, that has important semantic effects and interacts with many other phenomena, such as factivity and sentiment/polarity. Consider the example sentence below:
It is by no means ideal
Here, the phrase by no means functions as a so-called negation cue and it has the sentiment-bearing word ideal within its so-called scope, and thereby serves to reverse the overall polarity from positive to negative.
The task of negation resolution in NLP aims at detecting negation cues and the scope of these cues in natural language. Training and evaluating models for this requires annotated data, and the SANT project has recently released the first negation dataset for Norwegian; NoReCneg. An example of an annotated sentence is shown below. The texts are taken from the Norwegian Review Corpus – NoReC.
The focus of this thesis will be on the application of neural machine learning models to the task of negation resolution for Norwegian. We will experiment with different approaches and architectures, e.g., sequence-labelling approaches and graph-based approaches.