A key approach to carry out sentiment analysis, in particular, aspect-based sentiment analysis nowadays is using machine learning techniques that employ labeled datasets for training and tuning models to a specific task, thus making the development of relevant datasets the centerpiece of sentiment related research efforts. This paper presents a methodology for automated dataset construction, based on a gold standard annotated corpus of customer feedback on the services provided by private medical institutions. The annotation tagset comprises both tags for domain-specific concepts interpreted as opinion aspects and sentiment indicators. The proposed procedure for gold-corpus-based dataset creation includes: a) quasi-sentence segmentation of the customer review; b) token-for-token alignment of the dataset segments over the annotated corpus; c) aspect labeling of sentential units with N-dimensional binary vectors based on aspect tag extraction from the lexical units in the annotated corpus; d) sentiment labeling of the quasi-sentences based on transposing the sentiment-related tags of the corpus tokens into the values of a numerical scale. The developed dataset is a collection of tabular data that comprise quasi-sentences from customer reviews as elementary aspect-specific opinion-bearers, labeled each with a binary feature vector of aspects that allows for learning aspects as a multi-label classification problem, and positive and negative sentiment labels as floating-point numbers from the interval [0; 1], that provide for learning the opinion-bearing quasi-sentence sentiment as a regression problem.