Evidence Summary: Music Information Seeking Behaviour Poses Unique Challenges for the Design of Information Retrieval Systems

Objective – To better understand music information seeking behaviour in a real life situation and to create a taxonomy relating to this behaviour to facilitate better comparison of music information retrieval studies in the future. Design – Content analysis of natural language queries. Setting – Google Answers, a fee based online service. Subjects – 1,705 queries and their related answers and comments posted in the music category of the Google Answers website before April 27, 2005. Methods – A total of 2,208 queries were retrieved from the music category on the Google Answers service. Google Answers was a fee based service in which users posted questions and indicated what they were willing to pay to have them answered. The queries selected for this study were posted prior to April 27, 2005, over a year before the service was discontinued completely. Of the 2208 queries taken from the site, only 1,705 were classified as relevant to the question of music information seeking by the researcher. The off-topic queries were not included in the study. Each of the 1,705 queries was coded according to the needs expressed by the user and the information provided to assist researchers in answering the question. The initial coding framework used by the researcher was informed by previous studies of music information retrieval to facilitate comparison, but was expanded and revised to reflect the evidence itself. Only the questions themselves were subjected to this iterative coding process. The answers provided by the Google Answer researchers and online comments posted by other users were examined by the author, but not coded for inclusion in the study. User needs in the questions were coded for their form and topic. Each question was assigned at least one form and one topic. Form refers to the type of question being asked and consisted of the following 10 categories: identification, location, verification, recommendation, evaluation, ready reference, reproduction, description, research, and other. Reproduction in this context is defined as “questions asking for text” and referred most often to questions looking for song lyrics, while evaluation typically meant the user was seeking reviews of works (p. 1029). Sixteen question topics were outlined in the coding framework. They included lyrics, translation, meaning (i.e., of lyrics), score, work, version, recording (e.g., where is an album available for purchase), related work, genre, artist, publisher, instrument, statistics, background (e.g. definitions), resource (i.e. sources of music information) and other. The questions were also coded for their features or the information provided by the user. The final coding framework outlined 57 features, some of which were further subdivided by additional attributes. For example, a feature with attributes was title. The researcher further clarified the attribute of title by indicating whether the user mentioned the title of a musical work, recording, printed material or related work in their question. More than one feature could appear in a user query. Main Results – Overall, the most common questions posted on the Google Answers service relating to music involved identifying works or artists, finding recordings, or retrieving lyrics. The most popular query forms were identification (43.8%), location (33.3%), and reproduction (10.9%). The most common topics were work (49.1%), artist (36.4%), recording (16.7%), and lyrics (10.4%). The most common features provided by users in their posted questions were person name (53%), title (50.9%), date (45.6%), genre (37.2%), role (33.8%), and lyric (27.6%). The person name usually referred to an artist’s name (in 95.6% of cases) and title most often referred to the title of a musical work. Another feature that appeared in 25.6% of queries was place reference, almost half of which referred to the place where the user encountered the music they were enquiring about. While the coding framework eventually encompassed 57 different features, a small number of features dominated, with seven features used in over 25% of the queries posted and 33 features appearing in less than 10%. The seven most common features were person name, title, date, genre, role, lyric, and place reference. Lee categorized most of the queries as “known-item searches,” even though at times users provided incorrect information and many were looking for information about the musical item but not the item itself (p. 1035). Other interesting features identified by the author were the presence of “dormant searches,” long standing questions a user had about a musical item, sometimes for years, which were reawakened by hearing the song again or other events (p. 1037). Multiple versions of musical works and the provision of information gleaned third hand by users were also identified as complicating factors in correctly meeting musical information needs. Conclusion – While certain types of questions dominated among music queries posted on the Google Answers service, there were a wide variety of music information needs expressed by users. In some cases, the features provided by the user as clues to answering the query were very personal, and related to the context in which they encountered the work or the mood a particular work or artist evoked. Such circumstances are not currently or adequately covered by existing bibliographic record standards, which focus on qualities inherent in the music itself. The author suggests that user context should play a greater role in the testing and development of music information retrieval systems, although the instability and variability of this type of information is acknowledged. In some cases this context could apply to other works (film, television, etc.) in which a musical work is featured. Another potential implication for music information retrieval system development is a need to re-evaluate the terminology employed in testing to ensure that it is the language most often employed by users. For example, the 128 different terms used in this study to describe how a musical item made the user feel did not significantly overlap with terms employed in a previous music information retrieval task involving mood classification conducted through MIREX, the Music Information Retrieval Evaluation Exchange, in 2007. The author also argues that while most current music information retrieval testing is task-specific – e.g., how can a user search for a particular work by humming a few bars or searching for a work based on its genre, in real life, users come to their search with information that is not neatly parsed into separate tasks. The study affirms a need for systems that can combine tasks and/or consolidate the results of separate tasks for users.