LAUSR

Rationale refers to the reasoning and justification behind human decisions, opinions, and beliefs. In software engineering, rationale management focuses on capturing design and requirements decisions and on organizing and reusing project knowledge. This paper takes a different view on rationale written by users in online reviews. We studied 32,414 reviews for 52 software applications in the Amazon Store. Through a grounded theory approach and peer content analysis, we investigated how users argue and justify their decisions, e.g., about upgrading, installing, or switching software applications. We also studied the occurrence frequency of rationale concepts such as issues encountered or alternatives considered in the reviews and found that assessment criteria like performance, compatibility, and usability represent the most pervasive concept. We identified a moderate positive correlation between issues and criteria and furthermore assessed the distribution of rationale concepts with respect to rating and verbosity. We found that issues tend to appear more in lower star rated reviews, while criteria, alternatives, and justifications seem to appear more in three star rated reviews. Also, reviews reporting alternatives seem to be more verbose than reviews reporting criteria. A follow-up qualitative study of sub-concepts revealed, that users also report other alternatives (e.g., alternative software provider), criteria (e.g., cost), and decisions (e.g., on rating software). We then used the truth set of manually labeled review sentences to explore how accurately we can mine rationale concepts from the reviews. We evaluated the classification algorithms Naive Bayes, Support Vector Machine, Logistic Regression, Decision Tree, Gaussian Process, Random Forest, and Multilayer Perceptron Classifier using a baseline and random configuration. Support Vector Classifier, Naive Bayes, and Logistic Regression, trained on the review metadata, syntax tree of the review text, and influential terms, achieved a precision around 80% for predicting sentences with alternatives and decisions, with top recall values of 98%. On the review level, precision was up to 13% higher with recall values reaching 99%. Using only word features, we achieved in most cases the highest precision and highest recall respectively using the Random Forest and Naive Bayes algorithm. We discuss the findings and the rationale importance for supporting deliberation in user communities and synthesizing the reviews for developers.