Slide 1
Slide 1 text
Vowel Purity and Rhyme
Evidence in Old Chinese
Reconstruction
Johann-Mattis List (CRLAO)
Slide 2
Slide 2 text
Introduction
Slide 3
Slide 3 text
Rhyme Evidence in Old Chinese Reconstruction
● the morpheme-syllabic character of the Chinese writing
system does not give us many clues regarding the exact
pronunciation of Chinese during its oldest stages
● rhyme patterns in old Chinese poetry, like the Shījīng 詩經
(1050 - 600 BC), are therefore important for Old Chinese
reconstruction
● rhyme pattern analysis has a long tradition in Chinese
traditional scholarship, but it has never been thoroughly
systematized
Slide 4
Slide 4 text
Traditional Rhyme Pattern Analysis
● traditional rhyme pattern analysis (sīguàn shéngqiān fǎ 絲
貫繩牽法 follows a greedy approach
○ start from words which could be shown to rhyme with
each other in one poem
○ cluster words greedily into clusters by looking for
words which occur across poems
Slide 5
Slide 5 text
Traditional Rhyme Pattern Analysis
Slide 6
Slide 6 text
Traditional Rhyme Pattern Analysis
Slide 7
Slide 7 text
Traditional Rhyme Pattern Analysis
Slide 8
Slide 8 text
Traditional Rhyme Pattern Analysis
Slide 9
Slide 9 text
Traditional Rhyme Pattern Analysis
Slide 10
Slide 10 text
Traditional Rhyme Pattern Analysis
Slide 11
Slide 11 text
Traditional Rhyme Pattern Analysis
● unfortunately, the traditional rhyme analysis favors
lumping of rhyme categories over splitting
● no more than about 30 categories were identified by the
traditional rhyme analysis up to the middle of the 20th
century
● only later, Baxter’s (1992) hypotheses-testing approach to
quantitative rhyme data made it possible to postulate
more distinct (52) categories
Slide 12
Slide 12 text
Vowel Purity and Rhyme Evidence
Slide 13
Slide 13 text
Vowel Purity and Rhyme Evidence
● Ho (2016) criticises Baxter and Sagart’s (2014)
reconstruction of Old Chinese by pointing to many rhymes
in which words with different main vowels rhyme
● this principle, that says that old Chinese poetry was strictly
avoiding the rhyming of words with different vowels could
be called the principle of “vowel purity” in rhymes
Slide 14
Slide 14 text
Vowel Purity and Rhyme Evidence
● Ho’s (2016) argument rests on two fundamental
assumptions
a. Baxter and Sagart’s (2014) Old Chinese reconstruction
is in strong conflict with the principle of vowel purity
b. vowel purity was a key principle in Old Chinese
rhyming
Slide 15
Slide 15 text
Vowel Purity and Rhyme Evidence
● assumption b. is very difficult to check, and we find many
counter-examples both in Chinese rhyme traditions and in
a cross-linguistic comparison of rhyme traditions
● assumption a. can be easily checked, but unlike Ho
(2016), we need to check it quantitatively and
comparatively for different OC reconstruction systems
Slide 16
Slide 16 text
Vowel Purity and Rhyme Evidence
● Ho’s (2016) argument against Baxter and Sagart (2014):
○ lacks any concrete examples
○ confuses conflicts between traditional rhyme
categories and the rhyme categories by Baxter and
Sagart (2014) with actual conflicts with vowel purity
Slide 17
Slide 17 text
Vowel Purity and Rhyme Evidence
● Ho’s (2016) argument against Baxter and Sagart (2014):
○ lacks any concrete examples
○ confuses conflicts between traditional rhyme
categories and the rhyme categories by Baxter and
Sagart (2014) with actual conflicts with vowel purity
“It is a firm linguistic fact that rhyming should be based on identity of vowels.
Interpretation should, of course, be based upon facts. Facts precede and matter
more than interpretation.” (Ho 2016: 183)
Slide 18
Slide 18 text
Vowel Purity and Rhyme Evidence
● Ho’s (2016) argument against Baxter and Sagart (2014):
○ lacks any concrete examples
○ confuses conflicts between traditional rhyme
categories and the rhyme categories by Baxter and
Sagart (2014) with actual conflicts with vowel purity
“It is a firm linguistic fact that rhyming should be based on identity of vowels.
Interpretation should, of course, be based upon facts. Facts precede and matter
more than interpretation.” (Ho 2016: 183) → yes, let’s work with pure facts!
Slide 19
Slide 19 text
Evaluating Vowel Purity in
Reconstruction
Slide 20
Slide 20 text
Materials: Rhyme Network
The Shījīng Browser
● rhyme data from the Shījīng following Baxter (1992)
● digitized and converted to machine-readable format in List
(under review)
● data online available in form of a Shījīng Browser (http:
//digling.org/shijing/)
Slide 21
Slide 21 text
Materials: Rhyme Network
DEMO of
http://digling.org/shijing/
Slide 22
Slide 22 text
Materials: Rhyme Network
The Shījīng Rhyme Network
● a network of all words rhyming in the Shījīng (List under review)
● rhyme words are nodes in the networks (1996 nodes in total)
● links between nodes reflect instances in which two words rhyme in the
Shījīng according to Baxter’s (1992) analysis
● an automatic analysis of the patterns, in which community-detection
algorithms were used to search for potential rhyme groups is available
at http://digling.org/shijing/infomap.html
Slide 23
Slide 23 text
Materials: Rhyme Network
DEMO of
http://digling.org/shijing/infomap.html
Slide 24
Slide 24 text
Materials: Reconstruction Systems
● Baxter and Sagart (2014): online available for download
● Karlgren (1954): provided by Eastling (http://www.eastling.org)
● Wáng Lì (1980): provided by Eastling
● Pān Wúyùn (2000): provided by Eastling
● Zhèngzhāng (2003): provided by Eastling
● Starostin (1989): provided by Tower of Babel (http://starling.
rinet.ru)
● Li (1971): provided by Eastling
Slide 25
Slide 25 text
Materials: Reconstruction Systems
● Not all data is complete, since not all sources give
reconstructions for all characters in the Shījīng
● 1213 character readings occur in all seven datasets
● two analyses are carried out:
○ “complete coverage”: analysis for the character readings
which occur in all seven reconstruction systems
○ “partial coverage”: analysis for all character readings
available for a given reconstruction system
Slide 26
Slide 26 text
Materials: Reconstruction Systems
Slide 27
Slide 27 text
Methods: Testing Vowel Purity in Rhyme Networks
The Problem
● List (under review) uses community detection algorithms to
determine possible rhyme categories in the Shījīng and to
compare these with Old Chinese
● vowel purity, however, does not exclusively determine which
words rhyme with each other, since we know that other
aspects, like the coda, also contribute to rhyming
● on the other hand, vowel purity should restrict certain rhymes
Slide 28
Slide 28 text
Methods: Testing Vowel Purity in Rhyme Networks
The Problem
We search for a measure that reflects the tendency of vowel purity
in a network model. But vowel purity does only to a certain extent
coincide with rhyme categories. We thus do not search for a
measure that tells us something about the quality of communities
that we determine, but a measure that tells us to which degree the
topology of our network is in conflict with the characteristics of the
nodes.
Slide 29
Slide 29 text
Methods: Testing Vowel Purity in Rhyme Networks
Slide 30
Slide 30 text
Methods: Testing Vowel Purity in Rhyme Networks
Slide 31
Slide 31 text
Methods: Testing Vowel Purity in Rhyme Networks
Slide 32
Slide 32 text
Methods: Testing Vowel Purity in Rhyme Networks
Slide 33
Slide 33 text
Methods: Testing Vowel Purity in Rhyme Networks
Slide 34
Slide 34 text
Methods: Testing Vowel Purity in Rhyme Networks
Conductance as a measure of cluster purity?
● The conductance of a group of nodes in a network estimates the
degree of its isolation or fragmentation (Leskovec et al. 2008).
● List (under review) uses conductance to compare the purity of six-
vowel systems (Baxter and Sagart 2014) in contrast with Middle
Chinese vowel systems in the Shījīng rhyme network.
● But conductance has shortcomings when comparing different clusters
across graphs: When averaging conductance scores for each cluster
(vowel), systems with less vowels are favored.
Slide 35
Slide 35 text
Methods: Testing Vowel Purity in Rhyme Networks
Modularity as a measure for cluster purity?
● Modularity of a given set of clusters in a network is the fraction of the
edges within a cluster subtracted by the number of edges expected at
random. (Newman 2006)
● Modularity can be positive or negative, with positive values indicating
that communities are potentially present. (Newman 2006)
● Modularity suffers from a low resolution limits if networks become
large or nodes do not share many links.
Slide 36
Slide 36 text
Methods: Testing Vowel Purity in Rhyme Networks
Assortativity
● Assortativity (Newman 2003) tests whether nodes sharing
connections in a graph are also similar regarding other characteristics.
● Adapted to our rhyme network, this means that we test whether words
that rhyme in the Book of Odes share also the same vowel.
● This seems to be exactly what we are looking for: a measure for the
degree to which a given reconstruction system reflects the
assumption that words with identical vowels tend to rhyme.
Slide 37
Slide 37 text
Methods: Testing Vowel Purity in Rhyme Networks
Assortativity
high assortativity
Slide 38
Slide 38 text
Methods: Testing Vowel Purity in Rhyme Networks
Assortativity
low assortativity
Slide 39
Slide 39 text
Results
Slide 40
Slide 40 text
Results: General Remarks
Slide 41
Slide 41 text
Results: General Remarks
Slide 42
Slide 42 text
Results: General Remarks
Slide 43
Slide 43 text
Results: General Remarks
Slide 44
Slide 44 text
Results: General Remarks
Slide 45
Slide 45 text
Results: General Remarks
Coding Rhymes
by Vowel Quality:
■ a
■ e
■ i
■ o
■ u
■ ə
Slide 46
Slide 46 text
Results: General Remarks
Coding Rhymes
by Vowel Quality:
■ a
■ e
■ i
■ o
■ u
■ ə
Slide 47
Slide 47 text
Results: General Remarks
Coding Rhymes
by Vowel Quality:
■ a
■ e
■ i
■ o
■ u
■ ə
Slide 48
Slide 48 text
Results: Detailed Comparison
1213 Nodes 1471 - 1996 nodes
Slide 49
Slide 49 text
Results: Detailed Comparison
1213 Nodes 1471 - 1996 nodes
Slide 50
Slide 50 text
Results: Detailed Comparison
1213 Nodes 1471 - 1996 nodes
None of the systems shows a 100% vowel purity, but apparently, Baxter and Sagart (2014)
outperform all other reconstructions regarding vowel purity!
Slide 51
Slide 51 text
Conclusion and Discussion
Slide 52
Slide 52 text
Discussion
● the reconstruction by Baxter and Sagart (2014) corresponds
closer to the criterion of vowel purity than the other systems
compared
● the quantitative investigation shows that the critics by Ho (2016)
do not hold
● the rather high assortativity scores reported for almost all
reconstruction systems shows that vowel purity is a principle
that is reflected in all reconstructions (albeit with different rigor)
Slide 53
Slide 53 text
Discussion: Caveats
1. the comparison lacks full coverage for all reconstruction
systems, which may have influenced the results (although I
expect no larger differences)
2. errors both in the rhyme networks and the reconstruction
systems might have further influenced the results
3. we should be careful with the idea of vowel purity itself: it is by
no means proven that it was a driving principle for those who
created the poems in the Shījīng
4. we should ideally compare our study with an alternative rhyme
network, like the one by Wáng (1980)
Slide 54
Slide 54 text
Conclusion
Despite potential errors and further limits of the analysis, this study could
(hopefully) show that
● thorough quantitative comparison can give us new insights into our
problems in Old Chinese reconstruction
● instead of dismissing theories or reconstructions by spurious or non-
existing examples or assumptions, exhaustive evaluations give us a
fresh perspective on our problems
● In order to tackle our data-problems in the future, collaborative efforts
are required and people should try to share all their data as
transparently as possible
Slide 55
Slide 55 text
Many thanks to Philippe Lopez (Team Adaptation, Integration,
Reticulation, Evolution, UPMC) for providing invaluable help with
the network analysis, and to Laurent Sagart (CRLAO) for helpful
discussions and for providing the Old Chinese data!
Supplementary material with all code and all detailed
reconstructions available at:
https://gist.github.com/LinguList
Slide 56
Slide 56 text
Thanks for your attention!
Many thanks to Philippe Lopez (Team Adaptation, Integration,
Reticulation, Evolution, UPMC) for providing invaluable help with
the network analysis, and to Laurent Sagart (CRLAO) for helpful
discussions and for providing the Old Chinese data!
Supplementary material with all code and all detailed
reconstructions available at:
https://gist.github.com/LinguList