Advisory Board Online Conference 26 May to 7 June 2017

PI Egan wrote to project's Advisory Board:

<<
The project now has TEI-XML transcriptions of the 1608 Q1
and 1623 Folio texts of King Lear and is marking them
up to show the bits that scholars have traditionally
said are common to the two editions and the bits that
have traditionally been thought unique to one or other
edition. Attached is the crib-sheet that we use for this
purpose and any critiques you can make of that would be
very helpful.

But what I really want you for is to suggest experiment
we should do. The kinds of textual-comparison tests we
can do include:

* Measurement of the Shannon Entropy and Jensen-Shannon
Divergence of texts

* Word Adjacency Networks using Markov chains to store
the proximity values of 100+ function words found within
a text

* Nearest Shrunken Centroid applied to frequencies of
any words we care to count

* Edit Distance between texts

* Dynamic Time Warping to align to similar-but-not-identical
texts

By marking up Q1 and F, we have essentially four blocks
of writing to experiment with:

1) The Q1-only passages (hereafter Q1-unique)

2) The F-only passages (hereafter F-unique)

3) Q1 minus the Q1-only passages (hereafter Q1-common-with-F)

4) F minus the F-only passages (hereafter F-common-with-Q1)

To get us started, we thought we'd try to answer the
following questions:

Q1) Some basic stats. How long are the Q1-unique and F-unique
passages? How do their contributions break down by scene
or act-scene (as in can we graph where in the play--beginning,
middle, end--they fall)?

Q2) What differences are there between Q1-common-with-F and
F-common-with-Q1? Editorial judgement says that these bodies
of writing are essentially the same lines with small verbal
variants between them but no whole lines unique to either.
Do our tests substantiate that?

Q3) What differences if any are there between Q1-unique and
Q1-common-with-F? I'm thinking not only of differences in
choices of words within speeches, but also in kinds of speech
(prose versus verse) and who speaks. Can we say that certain
characters are over/under-represented in the Q1-unique material?
We are looking to see if Q1-unique seems to be 'of a piece'
with Q1-common-with-F (and indeed with F-common-with-Q1 if that
is not essentially the same as as Q1-common-with-F) or seems
different in some way.

Q4) What differences if any are there between F-unique and
F-common-with-Q1? I'm thinking of the same kind of questions
as in (2) above.

Q5) What differences are there between Q1-unique and F-unique?
In the one-text theory there was only ever one play and
Q1 and F are both imperfect records of it, each deviating
from the authoritative script only by errors of transmission.
If that is true, then Q1-unique and F-unique should be alike
despite being printed 15 years apart since both are parts
of the same lost original play that comprise the superset
of Q1-unique + F-unique + the lines in common to both
plays. Are we seeing that kind of wholeness across the
various parts?

Here's where you come in: please could you let me know
what you think of the above questions--be brutally honest
--and either suggest improvements to them and/or suggest
wholly new questions that haven't thought to. You should
also critique our approach at a more fundamental level
if you think we are going about things the wrong way. We
are keen to hear and take your advice.
>>

John Burrows replied:

<< 
These are good questions, I think, and well worth pressing
in the manner you suggest. Is there also a role for Hugh's
version of Zeta?

But I wonder  whether it is worth asking a more general
question: do the several relationships between Q and F
versions have common properties or are they all different
cases? Can one reasonably speak of bad quartos as a kind?

You may well think that  Lear alone may be enough for the
moment.
>>

PI Egan replied to John Burrows:

<<
Because our projects are so similar, Hugh and I are
coordinating our labours so as not to duplicate our
efforts except where we want one team to try to
replicate the findings of the other. For now,
Zeta is something Hugh's exploring and my team
is not.

> But I wonder whether it is worth asking a more general
> question: do the several relationships between Q and F
> versions have common properties or are they all different
> cases? Can one reasonably speak of bad quartos as a
> kind?

That's actually one of the fundamental research questions
for our project and we hope to get to that. Right now we're
groping our way forward with more local, single-play questions
in order to check that we've got our basic tools working.
>>

Doug Duhaime contributed:

<<
I'm wondering what might turn up if you complemented
some of these lexical questions with some related syntactic
and semantic questions. One could train a Markov model
on part of speech sequences, dependency tree structures,
or verb tenses, for instance, and compare distributions
in Q1 and F. In a related way, I thought one could use
"brown clustering" [1][2] or a word embedding model to
obtain vector representations of tokens from each of
the four corpora [3], then run analysis on those word
vectors to obtain distance measures. 

The basic idea with the latter is to build a term
cooccurrence matrix in which each linguistic "type"
(unique word string) in a target language is given
one row and one column. Then update the cell value
at each column, row position to the number of times
the word in the given column cooccurs with the word
in the given row (where cooccurrence means the words
appear within n words of each other in any document
within a corpus, and a corpus is the largest collection
of texts from a given time and place you can lay hands
on). This operation produces an n by n matrix, where n
is the number of linguistic types in the dataset. If
you then reduce the number of columns with a dimension
reduction technique such as non-negative matrix factorization,
you’ll have a matrix that’s n by p, where p is perhaps <= 200.
Then each row in the matrix represents a given linguistic
type with a p-dimensional vector. For each word, one can
then retrieve the corresponding vector, and those vectors
can be compared for continuous similarity, clustered, or
even added and subtracted for higher-order linguistic
analysis.

I've found these kinds of methods very helpful in capturing
more of the latent signal in language than raw tokens
provide, as they preserve similarities that are sometimes
lost when we treat tokens as discrete phenomena.

I also wanted to raise the question of ground truth. I
think it would be highly interesting if there were some
known instances of the single text theory outside of
Shakespeare that could be studied, measured, and used
to help inform the interpretation of the studies your
team will pursue on your corpora. If there were some 
oral traditions (or similar) for which the ur text and
their multiple printings were available, running analysis
on the similarities those descendants shared with the ur
text could potentially be insightful for the case of the
Q1 and F texts.

Figure 2.2 of Craig & Kinney's Shakespeare, Computers,
and the Mystery of Authorship stands out as a phenomenal
instance of leveraging ground truth to evaluate model
predictions, and I was wondering if there’s another
sort of ground-truth that could be used to evaluate
model predictions of the Q1 + F transmission case.
Whether Philaster provides an expected case that’s
analogous to the folio case I can’t say, but the
general thrust of my thought was to suggest that
it would be highly ideal to have some null hypothesis 
that could be tested with the analytic approaches to
be leveraged against the folio case.

[1] http://aclweb.org/anthology/J/J92/J92-4003.pdf

[2] http://curtis.ml.cmu.edu/w/courses/index.php/Brown_clustering

[3] https://nlp.stanford.edu/projects/glove/


>>

PI Egan replied to Doug Duhaime:

<<
[Regarding 'ground truth'] Perhaps sermons could
fulfil this role: we have cases in which they are
supposed to have been published from aural capturing
(by shorthand) during oral delivery and then are
published again by their authors in more correct
versions. But doesn't the aural/oral vector muddy the
waters here? Your first suggestion seems to be that we
want texts which were printed in multiple versions in
which oral garbling is not suspected, in order to be
parallel to Q1/F King Lear in which oral garbling is
generally not suspected. Perhaps Q1 (1620) and Q2
(1622)
>>

MacDonald P. Jackson contributed:

I can't guess how suited the kinds of tests you mention
are to answering the questions you list. I think that
with respect to King Lear the most interesting question
is Q4: Does the F-unique material differ in any way from
the F-common-with-Q1 material? Of course there may not
be enough F-unique material for the tests to provide a
definite answer. What we most want evidence about is
whether the F-unique material was always part of the
play or was added by Shakespeare several years later,
as Gary Taylor argued in Division of the Kingdoms, or
isn't by Shakespeare at all, as P. W. K. Stone claimed
(most implausibly in my opinion) in The Textual History
of ‘King Lear'. Are the tests capable of giving us an
answer? Whether they are or not, running the tests may
well reveal something of potential significance that we
haven't known before.

What I suspect they might be able to give an answer to
is the much broader question of whether the Shakespeare
quartos once labelled ‘bad' (and examined in Alfred Hart's
Stolne and Surreptitious Copies) form a distinct category
that differentiates them from all their ‘good' counterparts
(and from other plays of the period with sound texts). 
>>

PI Egan replied to MacDonald P. Jackson:

<<
Agreed, ["whether the F-unique material was always part
of the play or was added by Shakespeare several years
later"] that's a key question.

In principle, our tests are capable of that. One problem,
though, is that Q1 and F were printed 15 years apart during
a period when, we know, certain aspects of the language
were changing rapidly. We don't want to mistake mere
'modernization' of the spelling and inflectional kind
(doth>does, hath>has), which could have been applied by
by a scribe or a compositor, for more significant changes.
Unfortunately, in our base transcriptions the words are not
lemmatized so such discriminations are hard to make. Hugh is
working on markup to record words under their modern spelling
forms and to record the grammatical functions of homographs,
but his texts are not ready yet. We'll hoping to find out now
what can be done without such sophistications.

Deciding whether the 'bad' quartos do indeed form a 
group is definitely where we hope to be headed.
>>

Hugh Craig contributed:

<<
I wonder if your alignment work could be extended to identify
cases within the shared portions of *Lear* [= Q1-common-with-Fand
F-common-with-Q1] where aligned strings (punctuation, 1-gram,
n-gram, spelling, contraction etc) differ repeatedly between
the two versions?  This would extend what was done in the
Craig and Kinney chapter on *Lear* where we used a bag of
words approach and then looked at individual cases by hand.

Good data to collect would be instances where Q1 has the A form
and F has B, where Q1 has B and F has A, where Q1 and F both
have A, and where Q1 and F both have B.  What I mean by A
and B is things like:

A "that" [as in "he said that he would come"]  
B "zero that" [as in "he said he would come"]

A "upon" 
B "on"

A colon
B full stop

A "God's wounds"
B "Zounds"

If there is a consistent one-directional difference
representing a good proportion of the cases overall
this might look like human agency rather than textual
indeterminacy a la Lene Petersen, and then the next
question would be whether this is more likely
compositorial or authorial. 

On hath>has, doth>does etc, it would be a great project
to look at these in variant versions of the same texts
to see if and how much the later-printed versions
incorporate these changes--especially cases where no
authorial revision is suspected--with the idea of
getting a handle on whether this was a common
compositorial change.
>>