Syntactic Information as an Aid in Scope Detection

Correct scope detection is an important part in making consistent predictions for many NLP tasks, but it can be a difficult problem, both for newer models and for human annotators. This thesis explores how syntactic information can be utilised at different steps of the modelling process.

In fine-grained sentiment analysis, three components are usually identified: a polar expression expressing the actual polarity of what we call an opinion, a target, which is what the opinion is about, and a source or holder, which is the one holding the opinion. While polar expressions typically vary syntactically, targets and holders tend to be noun phrases. Some of the best models for Norwegian still struggle with boundaries, but some of this can also be attributed to inconsistencies. Earlier use of pre-annotation with syntactic information seem to improve annotators' syntactic judgement, and this project seeks to explore if parsers could be useful on the other side of the annotation task. The student would have the opportunity to both explore parsing as a step in modelling, but also work with inconsistencies in the original data, as well as becoming familiar with existing parsing systems for Norwegian.

Emneord: sentiment analysis, syntax, parsing, scope
Publisert 9. okt. 2023 14:53 - Sist endret 9. okt. 2023 14:53

Veileder(e)

Omfang (studiepoeng)

60