Product reviews available online contain information on users' interaction
with the product in the form of narratives. These narratives describe the
users' experiences with using the product in the wild, that is, in real usage
situations. The large amount of reviews available makes for a potentially
interesting source of information related to product use.
This thesis addresses the following research questions about the information
available in reviews:
1. How do users describe their interaction with products in narratives
such as product reviews?
2. Can such information be used to improve future versions of the prod-
uct?
The results show that users write about product use in terms related to
standard and popularly researched aspects of usability and user experience,
e.g. effiiency, effectiveness, enjoyment, frustration. The frequency with
which different aspects are depicted in reviews differs significantly between
product domains. We also find that reviews contain descriptions of more
persistent usability issues. I devise automatic methods for classifying sentences
with regard to dimensions of both usability and user experience and
usability problems and perform a linguistic analysis of the content.
To assist with the automatic classification tasks, I compared supervised
machine learning algorithms with semi-supervised learning (SSL) algorithms
for text classification on several standard corpora. This comparison shows
that support vector machines normally are the best choice for text classification.
I also find that traditional feature vectors consisting of counts of function
words can be used for identifying the author of a translated document but the
method can be augmented by adding semantic information from documents.
Overall the results presented in this thesis help clarify the role of reviews
in relation to understanding both users and different aspects of product use.