Year: 2025
Type: dissertation
Abstract: The thesis explores the relationship between measures of affect derived from linguistic and acoustic features of spontaneous Finnish speech in the FinnAffect corpus.A pipeline for lemmatization and sentiment score calculation of spontaneous Finnish speech transcriptions was developed.To this end, different lemmatization tools for colloquial Finnish were tested.The Trankit Pipeline for Finnish yielded the best lemmatization results for spontaneous speech transcriptions in a qualitative evaluation.Furthermore, a mapping from categorical sentiment scores based on individual words towards the valence scale was conducted to derive an estimate of the overall valence of an utterance.In a supervised classification experiment, predicting discrete valence labels from acoustic features and valence arousal scores obtained from ExhuBERT and FinnSentiment models, no systematic relationship between the linguistic and acoustic levels of spontaneous Finnish speech could be found.The code basis for this study is made publically available on Github to enfoster further endeavors in this area: https://github.com/BlindeEule/FiVal.
Author: Markl, Kerstin
Institutions: Tampere University, University of Tampere
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7132884482
