Sentiment and Semantic Analysis and Urban Quality Inference using Machine Learning Algorithms

Abstract

Qualitative interviews are conducted by researchers to gain a deeper understanding of people’s opinions and perceptions about a specific topic. The analysis of such textual data is an iterative process and often time-consuming. To help researchers obtain an overview of the data and improve their coding process, the objective of this thesis was to investigate how state-of-art techniques can be used to classify sentiment and semantic orientation from qualitative interviews transcribed in Swedish. The results demonstrate that the implemented deep learning techniques, BERT and NER, are a possible and promising solution to achieve the stated goal. For the sentiment analysis, the Swedish BERT model KB-BERT was used to perform a multi-class classification task on a text sentence level into the three different classes: positive, negative, and neutral. For the semantic analysis, NER and String Search were used to perform multi-label classification to match domain-related topics to the sentence. The models were trained and/or evaluated on partially annotated datasets. Nevertheless, an important factor to consider is that these deep learning models are heavily data-driven and would need accurately annotated domain-specific data to reveal their full potential.