Extracting aspects of online reviews: What can we make of text similarity?
Penghao Wang, Julia Taylor Rayz · 2017
This paper describes an experiment in extracting various aspects of restaurant reviews. The aspect extraction and clustering was accomplished using word2vec model, and the results were compared to previously used approaches. While the results show high precision and recall, the clusters do not always correspond to human intuition. Moreover, even when the similarity is high, it is very hard to find properties that would describe concepts within a given cluster. This suggests a need of a technique that takes into account human intuition and knowledge, not just a way of collecting and analyzing very large amounts of text.