Same same but different: Finding similar user feedback across multiple platforms and languages

Users submit feedback about the software they use through application distributions platforms, i.e., app stores, and social media. Previous research has found that this type of feedback contains valuable information for software evolution, such as bug reports, or feature requests. However, popular applications receive thousands of feedback entities per day, making their manual analysis unrealistic. In this work, we present an approach to automatically identify similar user feedback across different languages and platforms. At the core of the approach is a word aligner that aligns words based on their semantic similarity and the similarity of their local semantic contexts. Additionally, we make use of machine translation, sentiment analysis, and text classification, to extract the sentiment polarity and content nature of user feedback written in different languages. We use the results of these components to compute a similarity score between user feedback pairs. We evaluated our approach on user feedback entities written in four different languages, and retrieved from five different mobile applications obtained from four different app stores and social networking sites. The obtained results are encouraging. Compared to human assessment, the overall performance for monolingual user feedback pairs yielded a strong correlation of 0.79. For the crosslingual feedback pairs the correlation was also strong, with a value of 0.78.