Analysis of information features in natural language queries for music information retrieval: Use patterns and accuracy

A major issue in current music information retrieval (MIR) research is the lack of empirical studies of users and real-life music information seeking behavior. In particular, our poor understanding of real-life queries is an impediment to developing MIR systems that meet the needs of real users. This study aims, by an empirical investigation of real-life queries, to contribute to developing a theorized understanding of how people seek music information. This is crucial for informing the design of future MIR systems, especially the selection of potential access points, as well as establishing a set of test queries that reflect the real-life music information seeking behavior.;Queries are collected from an online reference website and coded using content analysis. The taxonomies of needs and information features are established by an iterative coding process and the intercoder reliability of the categories is tested by standard measures. The empirical associations between the needs and features are measured by the co-occurrences among them. Queries with known answers are qualitatively examined in order to determine the accuracy of selected information features provided by users and to gain insight into use of these features in music information queries.;This study found that (i) most of the queries analyzed are known-item searches, (ii) most contain a wide variety of kinds of information, (iii) much of this information is false, inaccurate, or uncertain, and (iv) despite these inaccuracies and uncertainties many queries are successful. A theory from pragmatics is suggested as a partial explanation for the unexpected success of inaccurate queries.;Based on this study some recommendations for improving MIR systems can be made: (i) incorporating user context in test queries, (ii) employing terms familiar to users in evaluation tasks, and (iii) combining multiple task results are recommended. Information about related multimedia works and applying attributive/referential readings of descriptions in IR may also help improve the current MIR systems.