Evaluation and comparison of natural language and graphical user interfaces in "query-by-impressions" scenes

Impression-based retrieval of multimedia content is the best way of finding items that suit the preferences, feelings, or mental states of users. This kind of system presents a relatively short set of words that denotes impressions and users estimate one or more words from this with a graphical user interface (GUI) to input into the system. With this approach, an increased number of words presented leads to increased costs for the user, in terms of the time and labor required to enter impressions. Conversely, a set of words that is too limited will often leave users unable to accurately represent their target impressions. We therefore compare the accuracy of retrieval and convenience of using an NLI (natural language interface) and GUI, which we introduced into an impression-based music-retrieval system. The results reveal their respective advantages and disadvantages, and suggest guidelines on the design of a better NLI.