In contrast to the most commonly used Sentiment Analysis Methods, which seek to detect the polarity of sentiments on a sentence and or document level, Aspect Based Sentiment Analysis follows a finegrained approach, where polarities of extracted terms referred to domain specific categories - referred as 'aspects' - are determined. Datasets created during the Semantic Evaluation Conferences in 2014 and 2016 for the purpose of Aspect-based Sentiment Analysis are widely used as benchmark datasets for evaluating different machine learning and deep learning approaches as well as competing against each other in terms of their performances in specific tasks, given by the evaluation of the datasets. My thesis aimes to evaluate different classification approaches with state of the art models within a distinct subtask of ABSA - Aspect Term Extraction. Aspect Term Extraction has the goal of extracting explicit and implicit Aspects expressed in a sentence and is a crucial step in ABSA before determining the polarities of extracted aspects. As my main contribution, i curated my own dataset, consisting of german reviews of three international airlines from TripAdvisor, which are then labelled in terms of aspects as implicit and explicit ones and its polarities (e.g. sentiments) by university students of Regensburg and by myself. Current SOTA (State-of-the-Art) Models such as Transformers (e.g. BERT) will be trained on the dataset and evaluated with respect to the approaches applied to classifying the aspects.
Information-Extraction of implicit and explicit Aspects for Aspect-Based Sentiment Analysis
Dataset Curation
Annotation Process
Evaluation (Annotation)
ML-Training