Article Preview
Top1. Introduction
Multimedia information retrieval (MIR), such as videos, musical content, animation, and so on, is ubiquitous nowadays. Search engines, be these for information or sales purposes, identify video content pertaining to a query and present results as ready to be consumed in their appropriate mode while musical content providers mine preferences through social networks and other sources in order to assist implicit musical queries leading to playlists.
Research related to MIR has long focused on multimedia’s content for producing representations on which to perform retrieval related tasks, such as similarity measurement. The advancement and widespread penetration of virtual social networks has provided another source of information that is contextual to the actual content and mostly refers to social networks users’ interaction with or related to the multimedia content. Contextual representations have been shown to significantly boost information retrieval related results in an array of scenarios (Melucci, 2008; Karydis, Kermanidis, Sioutas & Iliadis, 2013).
Despite the aforementioned focus on the content and context derived descriptors from multimedia data, metadata also allow for direct interpretation of their respective multimedia content (Hanjalic, Lienhart, Ma & Smith, 2008). Metadata descriptors, for all their shortcomings, when existing and accurate, offer a set of mostly predefined textual descriptors that allow for fast and relatively computational cheap information retrieval. Moreover, existing text information retrieval methods can be used, up to a degree of success with almost no adaptation, alleviating thus the need for customised methods for preliminary results.
Numerous approaches as to the schemata that best describe multimedia content exist (Smith & Schirling, 2006). In almost all approaches, the notion of a very short textual description (title) of the content as well as attribution of the content to its author/performer is common a phenomenon. These two attributes, the title and author, although not of the best discriminative capacity, exhibit adequate representative capability and are assigned to the content, by its author, ad hoc. Another approach introduces four general outputs for design science research: constructs, models, methods, and instantiations (March & Smith, 1995). Authors pointed out that design research could contribute to the applicability of Information System (IS) research by facilitating its application to better address the kinds of problems faced by IS practitioners.
Given multimedia’s content- and context- based MIR successful research results, the focus to metadata proposed herein may initially sound anachronistic. Nevertheless, this is far from the truth, as metadata-based MIR is still required for a plethora of research and industry related activities, such as: aggregating multimedia information from textual references (e.g. screen-scraping from html pages), measuring similarity when content is not available (e.g. client-side playlist editing without need to stream data or burden the server), traditional keyword search in search engines, merging results in meta-search engines that do not host content due to intellectual property issues, and many more.
Existing methodologies focusing on metadata-based MIR can be broadly separated into two classes based on whether use of supportive to the actual metadata information is used or not. This supportive information is similar to the aforementioned contextual (but not necessarily from social media), requires collection, is usually not objective and although it may enhance the metadata it can also introduce noise (Metzler, Dumais & Meek, 2007). This work focuses on solely the title and author metadata information of the multimedia content (i.e. no use of supportive/contextual information is done) in order to perform similarity measurement.