What is the ISMIR Cloud Browser?

It is a web-interface to facilitate content-based access to the cumulative ISMIR proceedings (from 2000 up to present). Its usage is very straightforward: When you search for a term, the web-interface will return a list of related terms and documents.

The terms and documents are ranked according to relevance (visually represented by the grey bars in front of each entries), most relevant first. Once you have a list of terms and documents, you can click any of them to use that entry as a new search query.

In addition, the document entries have links to their corresponding pdf's and bibtex entries.

If the proceedings are online anyway, why not use Google instead?

A major difference to Google's approach is that the estimated relevance of a document for a query term is not based directly on the occurrence of the query in the document, but rather on the distance between the query and the document in a semantic space.

Both documents and terms are represented as points in this space, so it is equally easy to use a document for a query, as it is to use a term.

How does it work?

The ISMIR Cloud Browser relies on a latent semantic analysis (LSA) [Landauer, Foltz, & Laham, 1998] of the proceedings. This process requires a collection of documents, and a list of terms that you are interested. In our case the term list is derived from the documents by automatic term extraction [Jacquemin, & Bourigault, 2003].

Counting the frequency of occurrence of each term in each document, we construct a term-document matrix. In this matrix, terms are represented as vectors in a document space, and documents as vectors in a term space. These spaces are not directly comparable. By factorizing the term-document matrix using non-negative matrix factorization, both terms and documents are represented as vectors representing points in a common subspace. The dimensionality of this subspace is a parameter value to be chosen. Ideally, each dimension in the subspace corresponds implicitly to a topic. This means that for a particular dimension, the terms with the highest activations on that dimension, tend to be semantically coherent.

For the cumulative ISMIR proceedings, an 80-dimensional LSA yields topics related to for example key finding, source separation, implementation frameworks, and lyrics. The top-20 terms for these example topics are displayed in the table below:

Topic 76Topic 66Topic 0Topic 14
1key finding source separation code lyric
2audio keyfinding separation cuda rhyme
3spiral time frequency d2k syllable
4chew separation performance platform song lyric
5izmirli spectrogram m2k phoneme
6spiral array stereo hardware rhyming
7pitch class mask architecture vowel
8krumhansl blind programming language pronunciation
9fuzzy analysis audio source separation chuck rap lyric
10profile voice estimate marsyas word
11tonality mixture execution alignment
12key separation result prototyping consonant
13minor key monaural environment misheard lyric
14key profile audio source processor stressed syllable
15pitchclass determination mixing execute oh
16probe tone db programmer text
17temperley source signal itinerary log odd
18probe nmf throughput unstressed syllable
19key finding result voice separation java acoustic distance
20symbolic key finding voice imirsel imperfect rhyme

As mentioned above, both terms and documents are represented points in the topic subspace, and can be thought of as activation patterns for the set of derived topics. The relevance between terms and documents can simply be computed as the cosine similarity between the activation patterns.

Why is my own publication not the most relevant document for my name?

Since relevance of a document for a term is computed by comparing topic activation patterns rather than simple occurrence of the term in the document, it is possible for documents to be judged as relevant for terms that do not occur in the document, but are still considered related. In many cases, this is an advantage, for example because it makes the system more robust against the misleading effects of synonymy and polysemy.

For the same reason, it may occur that if your name as an author is strongly associated with a particular topic, querying the system with your name yields papers from other authors related to that topic.

Why is term foo not being recognized?

The list of terms included in the ISMIR Cloud Browser is constructed using automatic term extraction. An initial list of candidate terms is built by extracting all noun phrases from the documents (using part-of-speech tagging). This list contains many false positives, including uninformative noun phrases (e.g. previous section, related work), pdf-to-text conversion artifacts, and terms that do not form part of the discourse, like fragments of mathematical formulae, tables and figures.

In order to get rid of these nuisance terms, we apply a combination of various heuristics (based on grammatical patterns, lists of stop words, dictionary filtering, and term statistics). However, it is very hard to get rid of nuisance terms while keeping all semantically relevant terms, especially domain specific acronyms and neologisms, which do not occur in regular dictionaries.

The filtering helps reduce the number of terms from roughly 330,000 to roughly 22,000, but as you will notice when you use the ISMIR Cloud Browser, it is far from perfect. Nevertheless, the system keeps track of unsuccessful queries, so that omissions detected through usage can be corrected in a later stage.

Why do I get errors when I try to include provided BibTeX entries in my LaTeX document?

That is most likely because of either one or both of the following reasons:

  • The BibTeX entries are encoded in UTF-8 format. In case they contain non-ascii characters, you need to tell LaTeX to handle that encoding by adding \usepackage[utf8]{inputenc} to the preamble of your document.
  • The note entry of the BibTeX entries contain the urls of the cited documents, wrapped in the LaTeX command \url, in order to have it typeset correctly, and make it appear as a link in your document. Unfortunately, the \url command is not part of the standard LaTeX installation; To use it you need to add \usepackage{hyperref} to the preamble of your document.

Alternatively, you can manually replace non-ascii characters by LaTeX representations of the character, and remove the \url command in the BibTeX entries to get rid of the errors.

More information

For more details, see Grachten et al. (2009) and the corresponding poster.

Acknowledgments

The 2009 version of the ISMIR Cloud Browser was created with the help of Tim Pohle and Markus Schedl. The ISMIR Cloud Browser is hosted at OFAI, the Austrian Research Institute for Artificial Intelligence (thanks to Dominik Schnitzer!). The project has been realized through the use of free software. Financial support to this project has come from FWF, the Austrian Science Fund, under grant number Z159 (the Wittgenstein-prize awarded to Gerhard Widmer).

References

  • Landauer, T. K., Foltz, P. W., & Laham, D. (1998). Introduction to Latent Semantic Analysis. Discourse Processes, 25, 259-284.
  • Jacquemin, C. & Bourigault, D. (2003). Term extraction and automatic indexing. In The Oxford Handbook of Computational Linguistics. Mitkov, R. (ed.). Oxford University Press.
  • Grachten, M., Schedl, M., Pohle, T., and Widmer, G. (2009). The ISMIR cloud: A decade of ISMIR conferences at your fingertips. In Proceedings of the Tenth International Society for Music Information Retrieval Conference. Kobe, Japan.