MIR - is it time to change our name?
The Music Information Retrieval Community is quite a diverse
collection of researchers. There are musicologists, librarians,
traditional NLP folk, plus the signal processing and machine learning
geeks. The ISMIR 2008 call for papers lists around 30 topic areas
including such diverse topics as compression, computational musicology,
music recommendation, score following and performance analysis. This
year's ISMIR is putting a 'singular premium' on submissions that span
two or more disciplines. Clearly topic diversity is important to our
research community.
I suggest that our the name of our
discipline 'Music Information Retrieval' may be too constraining,
and as such is reducing topic diversity. To me, the term 'music
information retrieval' constrains our discipline to tasks that extract
information from a body of music. The term seems to eliminate from
consideration a number of interesting topics including automatic
accompaniment, track alignment for playlisting, compression, automatic
annotation.
This sentiment was echoed at the recent meeting of NEMISIG. There were a number of researchers interested in tasks such as automatic scoring of personal photo and video streams, making new devices that allow people to directly affect the music that the are listening to, making new music from old (such as the jingler), reproducing performances of great (but dead) musicians. These problems are not strictly information retrieval.
If
we want to increase topic diversity in the MIR community perhaps it is
time to change our name to something that better describes the research
that is taking place in our community. I humbly
suggest that the term 'Music Intelligence Research' is a better
description of our research community.
Mostly Interdisciplinary Research? ;-)
Intelligent research sounds good though.
Posted by elias on February 06, 2008 at 09:14 AM EST #
Paul, you suggest that "the term 'Music Intelligence Research' is a better description of our research community." I strongly support the general idea. In fact, for three years in a row, I proposed to the ISMIR Steering Committee that we change the phrase to "Music Information Research". At our meeting in Vienna, I pushed it hard, but -- to put it mildly -- I got little support from the rest of the committee! But I'm no longer a member of the SC, just of the community... Yes, there are arguments against the change, but I thought (and still think) they're pretty weak. Of course either your form or mine would preserve the ISMIR acronym that has become so well-known.
--Don
Posted by Don Byrd on February 06, 2008 at 10:06 AM EST #
Don:
I like 'music information research' just as much as 'music intelligence research'. What were some of the arguments against the change? Was there concern that we may lose some of the library science folks, or perhaps that we would get swamped with synthesis technology? I'm very curious.
Posted by Paul on February 06, 2008 at 10:19 AM EST #
I remember Michael Fingerhut calling for similar name changes even as far back as 2001/2002. I think his proposal at the time was "Music Informatics" as opposed to "Music Information".
But looking at the wikipedia page for Information Retrieval, we see this definition (http://en.wikipedia.org/wiki/Information_retrieval):
"Information retrieval (IR) is the science of searching for information in documents, searching for documents themselves, searching for metadata which describe documents, or searching within databases"
So if you think about it, a lot of the subject areas that you mention are forms of (music) information retrieval. Music recommendation is the problem of finding songs that you would be interested in hearing, from a database of documents. Score following is the problem of searching for notes/chords within a single document, i.e. searching for where in a score the note that I am currently playing belongs. Performance analysis is the problem is searching for stylistic cues and variations within a single document.
Information retrieval, as a discipline, encompasses all these various areas.
The problem, IMHO, is not with the term "IR". It is with the public understanding of IR in the post-Google world. The mindshare that web search engines, let's say Google in particular, has overtaken means that most people have come to think of IR as having only the "search for documents in a collection" definition. But that just isn't true. "Information" does not just mean "documents", and "retrieval" does not just mean "web or large document collection". By doing automatic score transcription (audio to symbolic) one is actually "retrieving" note and chord and beat onset "information" from a song.
So frankly, I see no problem with the name. What I think we need to change is our understanding of what the name means. Information Retrieval means so much more than "search".
Posted by jeremy on February 07, 2008 at 04:39 PM EST #
Here is another definition, from the IST in the EU (http://cordis.europa.eu/ist/ka1/administrations/publications/glossary.htm):
"Information Retrieval: The science and practice of identification and efficient use of recorded data."
So things like annotation definitely also fit within this scope (Annotation = identification). And things like compression have always been a focus of (text) information retrieval research. Think about a classic like Ian Witten's "Managing Gigabytes: Compressing and Indexing Documents and Images". If you don't deal with information compression, the size of your indices are going to be unmanageable. So I see no reason why compression should not be part of music information retrieval, too. It will be a different type of compression. But that's just because the data type is different, not because "information retrieval" is any different.
I do understand the concerns being expressed here. But I want to offer a solution that returns the definition of "information retrieval" to its true historical glory :-) All these things (compression, score following, etc.) are embodiments of information retrieval. Web search has tended to obscure that fact.
Posted by jeremy on February 07, 2008 at 04:56 PM EST #
Sorry, just one more reference, then I'd be curious to hear what the rest of you think. Here is the call for papers for the 2008 SIGIR conference:
http://www.sigir2008.org/callforpapers.html
Check out a sample of the variety:
* IR Theory (e.g., formal models)
* IR Architectures (centralized, distributed, federated, peer-to-peer)
* Content representation and indexing
* Interactive IR
* Classification
* Machine learning for IR (e.g., clustering, learning to rank)
* Information Extraction for IR
* Privacy in IR
* Metadata-based IR
* Social tagging
* Image IR
* Video IR
* Question answering
* Filtering (e.g., routing, collaborative filtering, topic tracking)
* Summarization
* Text mining (including speech mining, bibliometrics)
* Genomic IR
* Other domain-specific IR (e.g., legal IR, IR for chemical structures)
Think about the wide variety of machine learning, statistical processing, NLP "intelligence", social intelligence, library knowledge, etc. that are required to solve all these Information Retrieval problems: Collaborative filtering, summarization, Question Answering, Topic Detection and Tracking (e.g. automated following of a news story as it develops over time), social tagging, peer-to-peer IR.
I mean, chemical IR is one of the topics! And in order to do chemical IR, you'll have to be able to do analyses of structures and patterns in molecules, bonding affinities, etc.
So if that sort of thing is still included within the domain of "Information Retrieval", I see no problem with continuing to call ISMIR work that does analysis of performer style "Music" Information Retrieval. Know what I mean?
Thoughts? Don? Anyone else?
Posted by jeremy on February 07, 2008 at 07:58 PM EST #
Jeremy:
All good points, and well argued. As you say, the real problem is that when people hear the term 'information retrieval' they think about certain types of apps only. Some people interpret MIR as applications that only consume music, they don't produce music.
I was a bit concerned to hear researchers at NEMISIG apologize about their projects because they were not real MIR - because their app produced music.
Perhaps, as you say, re-education is all that is needd
Posted by Paul on February 08, 2008 at 06:16 AM EST #
Jeremy proposed 'a solution that returns the definition of "information retrieval" to its true historical glory :-) All these things (compression, score following, etc.) are embodiments of information retrieval. Web search has tended to obscure that fact.' And Paul responded that Jeremy had a point, and commented that "Perhaps... re-education is all that is needed." Perhaps. But re-education campaigns of any kind are notoriously difficult and expensive; who's going to pay for this one?
This situation reminds of the way the general public has always applied the term "song" to musical works of all kinds -- symphonies, string quartets, whatever -- to the great annoyance of classical music lovers like myself. I think using "song" so broadly is worse than using "information retrieval" so narrowly. Regardless, I don't think there's much hope for re-education in either case. But this is different because we're not concerned with the general public, only people in the field, right? Wrong. I think my proposal the ISMIR Steering Committee rejected addresses that objection quite well; I'll post it here shortly.
Paul said "Some people interpret MIR as applications that only consume music, they don't produce music." Surely _most_ people interpret it that way -- and they're not likely to change as long as they're using Google to _retrieve_ information.
Posted by Don Byrd on February 08, 2008 at 11:27 AM EST #
Heh, Don, I actually remember having to be corrected by you more than once about the "song" terminology issue, in 1998/1999 when we first started working on this. At the time, we were working on the "search" type of information retrieval.
Since the field was so new, one of our problems then was coming up with terminology that appropriately captured both the music and the information retrieval aspects of what we were doing. What we needed was a term that expressed the granularity at which we were working. For text information retrieval, there was standard terminology: "document". A document is an atomic unit of retrieval, the basic level of information segmentation, the basic unit of indexing. We needed a similar term for music. However, with music, we had many different types. As you say, we had symphonies, quartets, cantatas, arias etc. in addition to all the 3-minute "songs" that Tim added by 2001/2002. Adding to that complexity was the fact that we weren't retrieving full symphonies; we were retrieving segmented symphonies, i.e. movements.
What word should we have used, to refer to all this music in the same way? What word is the equivalent of "document"? I think the word that we finally agreed upon was "piece". A symphony was a "piece" of music, a quartet is a "piece of music, a "song" is even a "piece" of music.
So why do I bring this up? Where am I going with this?
The point I am trying to make is that in those early days of Music IR, we weren't actually doing any "re-educating", as such. We were inventing new terminology. We were coming up with words that expressed both ideas from music as well as ideas from information retrieval. In that sense, when I began using the word "song", I was not actually using the <b>music</b> definition of song. I was using one possible <b>music information retrieval</b> definition, i.e. "musical document". Now I'm not suggesting we should go back to "song" -- you were clearly against it. I am only trying to point out that "song" meant something different in the MIR domain than in the music domain. "Song" meant "atomic retrieval unit" or "document". We were inventing new words. I mean, if we had not been inventing terminology, even the word "piece" (of music) has problems when applied to the MIR domain. Check out the WordNet definition of "piece":
'musical composition, opus, composition, piece, piece of music (a musical work that has been created) "the composition is written in four movements"'
http://wordnet.princeton.edu/perl/webwn?s=piece%20of%20music
From this reading, it seems like "piece" of music means an entire musical composition or work: the whole symphony. Not just one movement. But then that becomes problematic, because the atomic "documents" that we were searching in OMRAS 1 were not only pieces, but subpieces. We used movements of a symphony as "pieces". So again, "piece" is just as wrong as "song", if we believe that "music" terminology should be 1:1 mappable onto "music information retrieval" terminology.
I contrast this invention of new terminology with the proper contextualization of old terminology. I think in the case of (music) information retrieval, we are not inventing new terminology. We are inventing new subdomains of information retrieval. Information retrieval has always meant more than just web search. It's only in the last 10 years of this 40+ year field that people learned this new, Google-based definition -- and yet they learned it rather quickly. I think they people can re-learn what information retrieval really means, just as quickly. I don't actually think the re-education process will be that hard, because we are starting off with a definition that has historical roots. We're not actually changing anything, we're not actually trying to invent new terminology, like trying to decide whether "song" or "piece" should mean "atomic music document".
All we have to do is point out that things like TDT (topic detection and tracking) is a form of (text) information retrieval. If people understand that, which I think they should, then it becomes trivial to draw the parallel between TDT and score following. In TDT you follow a particular story over time; you want to make sure that your information need "syncs up" with the flow of news "documents", that you hit every documnt relevant to your need, even if the story starts to change as new information is revealed. With score following, you want to make sure that the notes you are currently expressing (as which Chris Raphael plays the oboe) "sync up" with the flow of notes and chords from the "piece" of music. If people understand that TDT is IR, then they should easily understand that score following is MIR.
Posted by jeremy on February 08, 2008 at 04:41 PM EST #
Don, you write:
"I think my proposal the ISMIR Steering Committee rejected addresses that objection quite well; I'll post it here shortly."
Please, do post it, because there is probably something that I am not understanding about what you are saying. I'd like to understand it better.
BTW, it's nice to have these old-time discussions with you again! :-)
Posted by jeremy on February 08, 2008 at 04:42 PM EST #
what about Music Information Revival? :-)
Posted by fabien on February 12, 2008 at 07:22 AM EST #
Ok, thanks for the discussion, Don.
I see where you are coming from with the name. Even though Information Retrieval does mean all these other things, someone might be put off, and not submit or investigate the conference further, because the current meaning of IR has changed to mean "to Google".
But still, I think there are enough clues in the title, even still. For example, "music" and "information" are still there. And that should be enough to hook an interested researcher or practitioner enough to look at the Call for Papers. And the CFP does go on to explain all the various areas that ISMIR covers.. the Chris-Raphaelian score following, the social tagging, the traditional searching, etc.
Posted by jeremy on February 13, 2008 at 01:40 PM EST #
Whether the name of the conference has enough clues to hook an "interested" researcher depends on the definition of "interested". I'm pretty sure that some people won't look into a conference on "music information _retrieval_" who would look into one on "music information _research_"; to me, the question is how many, and I tend to think quite a few. I'm so overloaded with information, I make an awful lot of snap decisions on what to follow up on, and I don't think I'm unusual that way.
Besides, even insiders get confused! Why else did Elias Pampalk (in his list of music IR dissertations) leave out so much of what ISMIR accepts? Elias, are you reading this? I suppose I should post this on _his_ blog :-) ...when I have time.
Posted by Don Byrd on February 13, 2008 at 04:22 PM EST #
> Elias, are you reading this?
Don, yes I've been carefully following this interesting discussion. If you have some dissertations I could add the the list I've been maintaining, please let me know. I'd be happy to add them.
Posted by elias on February 13, 2008 at 05:36 PM EST #
Yes, I would like to ask of Don which dissertations he is thinking of. Related to that, I would like to ask Paul if he could point to examples of work from the NEMISIG group that were worried about not being "true" MIR. Paul, in what areas were the concerns? Are they similar to the areas that Don will identify, for MIR dissertations? (No need to give names, just describe the broad types of work.)
Posted by jeremy on February 13, 2008 at 06:22 PM EST #
I can't remember if any of the NEMISIG attendees specifically expressed doubts about whether they were "real" music information retrieval. But, thinking back over the morning's presentations, it seems like most of the research groups had at least one project that wasn't information retrieval in the sense of finding documents from a database, which is the main connotation of the phrase when I hear it.
I can think of projects on transcription, games, accompaniment, interfaces for making music, TCNJ's interactive conductor simulation, synthesizing expressive music performances, music visualization, music and emotion, etc. Such projects already make up a large part of the MIR field and I think that the name chosen to represent the field should reflect this. I like music intelligence research and music informatics research as both of those are more inclusive of this part of the community.
Addressing Jeremy and Don's points about re-educating people, I think it's difficult to change the meaning or connotations of a name and generally what's necessary is the invention of a new name that doesn't have the same baggage.
Posted by mim on February 14, 2008 at 06:22 PM EST #
But we are not playing games with music just to play games, right? The game is not just a game. We are not going to sell it to Milton Bradley or to Hasbro, and people aren't just going to play it at parties. We are doing something like the Luis van Ahn ESP game to get better tags/labels out of people, am I correct? The games are designed to elicit some sort of human intelligence related to the music. And the reason we are trying to get better tags/labels/intelligence out of people is because we want to be able to use those labels to do music recommendation, music classification, or music search.
That's what information retrieval is! It is any method that supports the organization of information. Not just ranked list search. That includes indexing, information representation, compression, storage, visualization, similarity functions in support of score alignment, similarity functions in support of recommendation, analysis, modeling, etc. There is no name change needed. As long as it is in support of the goal of allowing us to get a better handle (better understanding/organization) on music information, then it is IR. So the game is not an end unto itself. It is the means to an "information retrieval" end.
Now, if on the other hand the ultimate goal is not to achieve some sort of better organization (tagging, visualization, search, recommendation etc.) of music, but rather to come up with games that people like to play, while socializing with each other, and no information is ever captured and reused, then I agree: We do need a name change.
But I don't think our scope is that broad.
Posted by jeremy on February 15, 2008 at 01:59 PM EST #
Let me just give one more example of the scope of information retrieval. CMU has one of the top IR research labs in the U.S. Here is the list of their research projects that fall under the Information Retrieval category:
http://www.lti.cs.cmu.edu/Research/projects.html#ir
And many of these projects have analogues in the music IR domain:
Adaptive information filtering -> music recommendation
Distributed IR -> Playlist sharing
Email classification -> Music genre classification
eRulemaking -> Analysis and visualization of music structure (either on a per song or a per user collection basis)
Briefing Assistant -> Shenkerian analysis in support of music structure understanding
I won't go on. But look at the remaining CMU projects: Javelin (question answering), REAP, and Utility-based information distillation, and you'll see how wide the definition of Information Retrieval really is.
It is not just finding documents in a collection.
Posted by jeremy on February 15, 2008 at 02:32 PM EST #