In recent years, the challenging natural language processing  (NLP) task of aspect-based sentiment analysis (ABSA) as a target-aware sentiment classification problem has gained increasing traction in artificial intelligence research. Target-aware sentiment analyses enable finely grained sentiment classification by identifying the polarity of sentiments towards specific aspect terms such as companies, individuals, commodities, and currency pairs in unstructured, opinionated text such as financial news headlines. Sentiment expressions vary across different domains, which means that most relevant research relies heavily on domain-specific data sets or entity lists. Creating these data sets is time-consuming and requires domain knowledge. Even with sufficient domain knowledge, manual labelling of the data is often subject to scrutiny due to the biases of the individuals involved. In this paper, we propose a multi-step NLP algorithm for leveraging an optimised version of Bidirectional Encoder Representations from Transformers (BERT) in the aforementioned context. Unlike in most existing work, our approach does not focus on the sentiment classification sub-step of ABSA. Instead, emphasis is placed on the equally important aspect extraction sub-step by framing opinion target extraction (OTE) as a segment classification sequence labelling task. Our approach involves (a) the development of an initial model utilising publicly available labelled data and an optimised BERT-based model, and (b) an improved model refined by a sample of manually labelled financial news headlines validated by the initial model. We demonstrate that this approach is capable of efficiently extracting multiple opinion targets from financial news without extensive manual labelling or reliance on annotated data sets and entity lists.