Welcome

Welcome to the Corpus Studies workshop for the SMT Jazz Interest Group (2019).

Today we will:

• Discuss how corpus studies might be useful for your research.
• Discuss the constraints and issues related to corpus studies.
• Look at the essential tools and datasets.

Installation

• You might want to consider installing git if you don't already have it. There's a nice guide on how to do this on the Humdrum website.

Some Reading

Broze, Y., & Shanahan, D. (2013). Diachronic Changes in Jazz Harmony A Cognitive Perspective. Music Perception: An Interdisciplinary Journal, 31(1), 32–45.

Katz, J. (2017). Harmonic Syntax of the Twelve-Bar Blues Form. Music Perception, 35(2), 165–192.

Norgaard, M. (2014). How Jazz Musicians Improvise: The Central Role of Auditory and Motor Patterns. Music Perception: An Interdisciplinary Journal, 31(3), 271–287.

Salley, K., & Shanahan, D. (2016). Phrase Rhythm in Standard Jazz Repertoire: A Taxonomy and Corpus Study. Journal of Jazz Studies, 11(1).

Shanahan, D., & Broze, Y. (2012). A diachronic analysis of harmonic schemata in jazz. Proc. 12th Int. Conf. on Music Perception and Cognition and the 8th Triennial Conf. of the European Society for the Cognitive Sciences of Music, Thessaloniki, Greece, 23, 909–917.

Introduction

Rather than beginning on a typically optimistic note, let's begin by talking about what we won't be able to do in this brief, hour-long workshop:

• we won't be able to learn a programming language in two hours
• we won't even be able to learn about a single toolkit terribly well (be it music21, Humdrum, or anything else you might be interested in using).
• we won't be able to have a complete grasp on a broad interdisciplinary field that encompasses computer science, psychology, machine learning, musicology, music theory, statistics, and data visualization (among others!).

But all is not lost! Hopefully in this all too brief window we will:

• formulate a research question
• build a corpus that will try to get at aforementioned research question
• begin to tackle the research question in either an empirical or a descriptive fashion (although these two terms aren't necessarily diametrically opposed)
• TBD
• world domination.

Some questions we will be addressing today

• How might "distant readings" be employed in our research?
• What does it mean to operationalize something, and how skeptical should we be of the operationizations of others?
• What can we learn from global variables?

What is a corpus study?

For the purposes of today's workshop, we might think of a corpus study as a "distant reading", defined as:

• "a form of analysis that focuses on larger units and fewer elements in order to reveal patterns and interconnections through shapes, relations, models, and structures." (Burdick, et al., 2012)
• "a term that is specifically arrayed against the deep hermeneutics of extracting meaning from a text through ever-closer, microscopic readings." (Burdick, et al., 2012)
• "not just a “digitization” or “quickener” of classic humanities methodologies. It is, rather, a new way of doing research wherein computational methods allow for novel sets of questions to be posed about the history of ideas, language use, cultural values and their dissemination, and the processes by which culture is made" (ibid.)

For the purposes of today's workshop, let's define a corpus study as: what we do when we look at the relationships of music from a bird's-eye view.

Corpus Studies without Computers

Knud Jeppesen (1922, 1927) approached Palestrina's work in a "strictly scientific spirit".

• Work was done completely by hand.
• Toward the end of the dissertation, he a melodic idea and then cites the 143 spots in Palestrina's output.
• Elsewhere, he cites a cambiata figure and then lists 300 or more exemplars.

• Gjerdingen's Work in Classic Turn of Phrase
• Byros's (2009) dissertation on the le-sol-fi-sol motive in Beethoven.

Could we argue that Caplin's Classical Form and Hepokoski and Darcy's Elements of Sonata Theory are also corpus studies?

Their theories are constructed from the analysis of large collections of pieces. Is that a corpus analysis?

Why Corpus Studies?

• Analyzing the relationship of pieces to the larger whole allows for us to better understand relationships and context.
• All the cool kids are doing it.

What Data is Available?

• The iRealB Lead Sheet Corpus (see Shanahan and Broze, 2012; Broze and Shanahan, 2013)

• The Omnibook Corpus (see Baker, Shanahan, Rosado, and Shanahan, 2016).

• Audio data is available via the Spotify API and other services.

Audio Data vs. Symbolic Data

Symbolic Data is data encoded from a score. This can mean data encoded into musicXML, MEI, Kern, or MIDI.

Audio Data is data taken directly from a recording. It can be taken from something like Spotify (with the Spotify API) or it can be extracted with a tool such as Sonic Visualiser.

Symbolic Data

• PRO: Allows for the analysis of specific musical features, including scale-degree information, harmonic information, and meter.
• CON: Laborious to encode. It has to be done manually, meaning that your corpus will likely be a bit smaller.

Audio Data

• PRO: Most often extracted automatically, meaning you can get a lot more data.
• NEUTRAL: The "musical object" is not the score, but the performance.
• CON: Data extracted can be somewhat unreliable (beat finding, for example, is problematic; harmonies are very problematic), and often the metrics are somewhat difficult to decipher, and the mechanics are often kept a trade secret (what does danceability mean? Acousticness?).

Moretti on Operationalizing

• Operationalizing is "a process which is absolutely central to the new field of computational criticism, or, as it has come to be called, of the digital humanities." (p.1)
• "A theory-driven, data-rich research program has become imaginable, bent on testing, and, when needed, falsifying the received knowledge of literary study. Of this enterprise, operationalizing, will be the central ingredient." (p.13, emphasis added)

Falisifiability

As humans, we are extremely good at coming up with narratives and reasons for why something might be the case.

• Birds of a feather flock together.
• Opposites attract.

It's worth making the distinction between:

• exploratory analysis
• hypothesis-driven analysis

While these exploratory studies can be useful, at some point we need to come up with a way of minimizing our own ability to construct a narrative.

This is where a falsifiable hypothesis comes into play.

The most important thing to keep in mind is:

• What would it look like for me to be wrong?

One way to do this would be to construct an a priori hypothesis. Using inferential statistics, you could then test this hypothesis against a null hypothesis (in which there would be no significant difference between something observed and chance, for example.)

• With descriptive statistics, we describe a sample of a population, makes no broad claims about how something relates to broader populations.
• With inferential statistics, we are using the sample we have to infer conclusions about the broader sample.

Sometimes these lines can be a little blurry, at first. For example, we can have a directionless hypothesis, such as "these things are related", but not necessarily saying if they're positively or negatively correlated.

Final thoughts

• What questions might we have about a corpus?
• How might you go about asking a computer to find these things?