| ACE XML |
|---|
Update
This page describes ACE XML Version 1.1. This version is currently supported by all of the jMIR software components.
Work is currently underway to replace ACE XML 1.1 with the redesigned and significantly more expressive ACE XML 2.0. The ACE XML 2.0 specification is now complete, but the associated software is still under development. The 1.1 version of ACE XML described on this page will be deprecated once ACE XML 2.0 compatibility is incorporated into the jMIR components, although the jMIR components will remain backwards compatible with ACE XML 1.1.
Details on ACE XML 2.0 are available on the ACE XML 2.0 Development Page.
Overview
ACE XML is a set of file formats developed to enable communication between the various jMIR software components. These file formats have been designed to be very flexible and expressive, and it is hoped that they will eventually be adopted beyond the limited scope of jMIR as a multi-purpose standardized format by the music information retrieval (MIR) research community.
Aside from ACE XML, there is currently no unified standardized general-purpose format for storing specialized MIR-related information such as class ontologies, feature values and ground truth labels. This has resulted in the emergence of the Weka ARFF format as a de facto standard. Although ARFF and Weka in general are powerful tools, they have not been designed with the specific needs of music in mind, and thus have a number of important limitations when applied to MIR research.
Summary of Advantages of ACE XML over Weka ARFF and Other General Formats with Regard to MIR Research
As can be expected from any truly multi-purpose framework, Weka ARFF has a number of limitations with respected to the particularities of music and MIR. Although these problems are in no way limited only to Weka, which is in general an exceptionally good system, they can perhaps best be illustrated via an analysis of the Weka ARFF file format which is used to communicate feature values and class labels to Weka. The following points demonstrate the limitations of the Weka ARFF format that are met by ACE XML:
- There is no good way to assign more than one class to a given instance in an ARFF file, something that one might often wish to do in MIR tasks such as genre classification. There are awkward workarounds, such as breaking one multi-class problem into many binary classification problems so that there is a separate ARFF file for every class, but this is obviously a non-optimal approach. ACE XML provides the ability to assign multiple classes to individual instances.
- ARFF files do not allow any labeling or structuring of instances. Each instance is stored only as a collection of feature values and a class identifier, with no identifying metadata or sub-segmentations. This is problematic when dealing with music where, for example, it is often appropriate to extract features over a number of possibly overlapping related windows. Furthermore, some features may be extracted for each window, some only for some windows and some only for a recording as a whole. ARFF files provide no way of associating the features of a window with the recording that it comes from, nor do they provide any means of identifying recordings, storing additional metadata relating to them or of storing time stamps associated with each window. ACE XML, in contrast, allows all of these things.
- ARFF files do not permit any logical grouping of features. Each feature is treated as an independent quantity with no structural relation to any other features. This makes it difficult to fully capitalize on the advantages offered by multi-dimensional features. ACE XML allows logical grouping of features.
- There is no way of imposing a structure on class labels in ARFF files. One often encounters hierarchical or more sophisticated ontological structures in music (e.g., structural analysis or genre classification), and it can be very profitable to take advantage of this using appropriately structured classifier ensembles or weighted misclassification penalties during training. ACE XML allows the specification of hierarchical class ontologies.
Details of ACE XML
An important priority when developing a feature file format is to enforce a clear separation between the feature extraction and classification tasks, as particular researchers may have reasons for using particular feature extractors or particular classification systems. A successful file format should therefore make it possible to use any feature extractor to communicate any features of any type to any classification system. This portability makes it possible to use features generated with different extractors with the same classification system, or a given set of extracted features with multiple classification systems.
The reusability of files is another important consideration. For example, it could be useful to use the same set of extracted features for a variety of tasks, such as genre classification as well as artist identification. Similarly, it could be convenient to reuse the same model classifications with different sets of features. For example, one could classify a given corpus of audio recordings and then later perform the same task on symbolic recordings of the same corpus using the same model classifications. Unfortunately, most current feature file formats merge feature values and model classifications, making this kind of reusability difficult.
The use of two separate types files was therefore adopted in ACE XML for what is traditionally contained in one file, namely one file for storing feature values and another for storing model classifications. Unique keys such as file names can be used to merge the two files. The model classification file can be omitted when using unsupervised learning or classifying unknown patterns.
ACE XML also includes an additional optional file format for specifying taxonomical class structures. This enables one to specify the relationships between classes, information that can be very useful for tasks such as hierarchical classification. This file can be omitted if only flat classification is to be used. There are plans to expand ACE XML to include more general ontologies.
The final optional file format included in the ACE XML specification stores metadata about features, such as basic descriptions or details about the cardinality of multi-dimensional features. Although not strictly necessary, such a file helps solidify the potential for full independence between feature extractors and classifiers. A researcher with a classifier could be e-mailed a feature values file and a feature definitions file by other researchers, for example, and would need no additional information about the feature extractor used or the features it extracted.
<!ELEMENT feature_vector_file (comments, data_set+)>
<!ELEMENT comments (#PCDATA)>
<!ELEMENT data_set (data_set_id, section*, feature*)>
<!ELEMENT data_set_id (#PCDATA)>
<!ELEMENT section (feature+)>
<!ATTLIST section start CDATA "" stop CDATA "">
<!ELEMENT feature (name, v+)>
<!ELEMENT name (#PCDATA)>
<!ELEMENT v (#PCDATA)>Figure 1: XML DTD of the ACE XML file format for storing feature values.
The explicit XML Document Type Definitions (DTD’s) of the four proposed ACE XML formats are shown in Figures 1 to 4. It can be seen from Figure 1 that features may be stored for overall instances, called data sets, which may or may not have sub-sections. This can correspond to a recording and its windows, for example. Each sub-section has its own features, and each data set may have overall features as well. Each sub-section may have start and stop stamps in order to indicate what portion of the data set it corresponds to. This makes it possible to have windows of arbitrary and varying sizes that can overlap. Each feature has a name identifying it, which makes it possible to omit features from some data sets or sub-sections if appropriate. Each feature may also have one or more values (denoted by the <v> element) in order to permit multi-dimensional features.
<!ELEMENT classifications_file(comments,data_set+)>
<!ELEMENT comments (#PCDATA)>
<!ELEMENT data_set (data_set_id, misc_info*, role?, classification)>
<!ELEMENT data_set_id (#PCDATA)>
<!ELEMENT misc_info (#PCDATA)>
<!ATTLIST misc_info info_type CDATA "">
<!ELEMENT role (#PCDATA)>
<!ELEMENT classification (section*, class*)>
<!ELEMENT section (start, stop, class+)>
<!ELEMENT class (#PCDATA)>
<!ELEMENT start (#PCDATA)>
<!ELEMENT stop (#PCDATA)>Figure 2: XML DTD of the proposed ACE XML file format for storing class labels.
Figure 2 shows the DTD for storing class labels, be they model labels or classification results. Each data set may have optional metadata associated with it. Each data set can be broken into potentially overlapping sub-sections if desired, and each sub-section can be assigned one or more classes. Each data set may be assigned one or more overall classes as well. Each sub-section is given start and stop stamps to show the region of influence of particular classes.
<!ELEMENT taxonomy_file (comments, parent_class+)>
<!ELEMENT comments (#PCDATA)>
<!ELEMENT parent_class (class_name, sub_class*)>
<!ELEMENT class_name (#PCDATA)>
<!ELEMENT sub_class (class_name, sub_class*)>Figure 3: XML DTD of the optional ACE XML file format for storing class taxonomies.
The DTD for the optional taxonomy format is shown in Figure 3. This format allows the representation of hierarchically structured taxonomies of arbitrary depth.
<!ELEMENT feature_key_file (comments, feature+)>
<!ELEMENT comments (#PCDATA)>
<!ELEMENT feature (name, description?, is_sequential, parallel_dimensions)>
<!ELEMENT name (#PCDATA)>
<!ELEMENT description (#PCDATA)>
<!ELEMENT is_sequential (#PCDATA)>
<!ELEMENT parallel_dimensions (#PCDATA)>Figure 4: XML DTD of the optional ACE XML file format for storing feature definitions.
The final optional file format, for storing feature definitions, is shown in Figure 4. This format enables one to store the name of each possible feature, a description of it, whether or not it can be applied to sub-sections or to overall data sets only and how many dimensions it has.
Sample ACE XML Files
Related Publications
McKay, C., R. Fiebrink, D. McEnnis, B. Li, and I. Fujinaga. 2005. ACE: A framework for optimizing music classification. Proceedings of the International Conference on Music Information Retrieval. 42–9.
Questions and Comments