Network Approaches to Old Chinese Reconstruction Johann-Mattis List Department of Linguistic and Cultural Evolution Max Planck Institute for the Science of Human History Jena 2017/06/30 1 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Overview long tradition of linguistic reconstruction in China (starting with Chén Dì 陳第, 1541 – 1606), 3 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Overview long tradition of linguistic reconstruction in China (starting with Chén Dì 陳第, 1541 – 1606), breakthrough in the early 20th century with Karlgren’s reconstructions and impressive work by Wáng Lí 王力 (1980) and Li Fang-kuei 李方桂 (1971), 3 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Overview long tradition of linguistic reconstruction in China (starting with Chén Dì 陳第, 1541 – 1606), breakthrough in the early 20th century with Karlgren’s reconstructions and impressive work by Wáng Lí 王力 (1980) and Li Fang-kuei 李方桂 (1971), since then more and more improved concrete reconstructions of Old Chinese phonology, 3 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Overview long tradition of linguistic reconstruction in China (starting with Chén Dì 陳第, 1541 – 1606), breakthrough in the early 20th century with Karlgren’s reconstructions and impressive work by Wáng Lí 王力 (1980) and Li Fang-kuei 李方桂 (1971), since then more and more improved concrete reconstructions of Old Chinese phonology, another breakthrough in the 1980s, when Baxter (1992), Starostin (1989), and Zhèngzhāng Shàngfāng (see Zhèngzhāng 2003) presented reconstructions in which they independently proposed several similar features (notably six vowels) 3 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Overview long tradition of linguistic reconstruction in China (starting with Chén Dì 陳第, 1541 – 1606), breakthrough in the early 20th century with Karlgren’s reconstructions and impressive work by Wáng Lí 王力 (1980) and Li Fang-kuei 李方桂 (1971), since then more and more improved concrete reconstructions of Old Chinese phonology, another breakthrough in the 1980s, when Baxter (1992), Starostin (1989), and Zhèngzhāng Shàngfāng (see Zhèngzhāng 2003) presented reconstructions in which they independently proposed several similar features (notably six vowels) current state-of-the-art is the reconstruction by Baxter & Sagart (2014) 3 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Problems data and judgments are often not made publicly available difficult to compare different judgments without additional data 4 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Problems data and judgments are often not made publicly available difficult to compare different judgments without additional data no large-scale comparisons of different types of evidence and how scholars interpret 4 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Problems data and judgments are often not made publicly available difficult to compare different judgments without additional data no large-scale comparisons of different types of evidence and how scholars interpret although the research is based on very large amounts of data, those data are usually not accessible in a digital format, which makes it difficult for scholars to evaluate and investigate a given reconstruction system, not to speak of comparing different ones 4 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Types of Evidence rhymes in ancient poems structure of ancient characters reconstruction of Middle Chinese via fǎnqiè readings, rhyme books, and ancient character descriptions (dúruò etc.) 5 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Types of Evidence rhymes in ancient poems structure of ancient characters reconstruction of Middle Chinese via fǎnqiè readings, rhyme books, and ancient character descriptions (dúruò etc.) Sino-Xenic readings 5 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Types of Evidence rhymes in ancient poems structure of ancient characters reconstruction of Middle Chinese via fǎnqiè readings, rhyme books, and ancient character descriptions (dúruò etc.) Sino-Xenic readings ... 5 / 35
Introduction Old Chinese Reconstruction Old Chinese Reconstruction: Types of Evidence rhymes in ancient poems structure of ancient characters reconstruction of Middle Chinese via fǎnqiè readings, rhyme books, and ancient character descriptions (dúruò etc.) Sino-Xenic readings ... Interestingly, many of the different types of evidence have been investigated with help of rudimentary network approaches by classical Chinese scholars in the past. 5 / 35
Introduction Networks Networks: Examples Many structures in daily life and science can be modeled as networks: social networks: nodes are persons, edges are relations between persons (e.g., friendship on FaceBook, etc.), 7 / 35
Introduction Networks Networks: Examples Many structures in daily life and science can be modeled as networks: social networks: nodes are persons, edges are relations between persons (e.g., friendship on FaceBook, etc.), phylogenetic networks: nodes are languages or dialect varieties, edges represent genetic closeness, 7 / 35
Introduction Networks Networks: Examples Many structures in daily life and science can be modeled as networks: social networks: nodes are persons, edges are relations between persons (e.g., friendship on FaceBook, etc.), phylogenetic networks: nodes are languages or dialect varieties, edges represent genetic closeness, network of sound change patterns: nodes are sounds, directed edges represent likelihood of sound change during language evolution, 7 / 35
Introduction Networks Networks: Examples Many structures in daily life and science can be modeled as networks: social networks: nodes are persons, edges are relations between persons (e.g., friendship on FaceBook, etc.), phylogenetic networks: nodes are languages or dialect varieties, edges represent genetic closeness, network of sound change patterns: nodes are sounds, directed edges represent likelihood of sound change during language evolution, ... 7 / 35
Introduction Networks Networks: Approaches identify clusters (communities) of nodes in the network check externally proposed groupings against grouping emerging from the network itself 8 / 35
Introduction Networks Networks: Approaches identify clusters (communities) of nodes in the network check externally proposed groupings against grouping emerging from the network itself investigate the general dynamics of the network 8 / 35
Introduction Data Data: Overview data should be easily accessible data should be transparent (easy to comprehend for experts) data should be human- and machine-readable (to allow for error-checking and automatic analysis as well as qualitative analysis 10 / 35
Introduction Data Data: Overview data should be easily accessible data should be transparent (easy to comprehend for experts) data should be human- and machine-readable (to allow for error-checking and automatic analysis as well as qualitative analysis data should be shared immediately with publications if publications rely on the data 10 / 35
Introduction Data Data: Increasing Comparability simple formats should be preferred over complicated ones tabular data can be edited in Excel but should be shared in form of CSV files 11 / 35
Introduction Data Data: Increasing Comparability simple formats should be preferred over complicated ones tabular data can be edited in Excel but should be shared in form of CSV files if data structures cannot be immediately understood, authors should add metadata to describe them 11 / 35
Introduction Data Data: Increasing Comparability simple formats should be preferred over complicated ones tabular data can be edited in Excel but should be shared in form of CSV files if data structures cannot be immediately understood, authors should add metadata to describe them data can be easily hosted on scientific repositories like Zenodo (zenodo.org), and curation is simple with help of services such as GitHub or GitBucket 11 / 35
Introduction Data Data: The CHIP Initiative (with N. W. Hill) try to digitize rhyme judgments of different authors (like Baxter’s 1992 Shījīng readings, Wáng’s 1980 Shījīng readings, etc.) make the data comparable, both for humans and machines 12 / 35
Introduction Data Data: The CHIP Initiative (with N. W. Hill) try to digitize rhyme judgments of different authors (like Baxter’s 1992 Shījīng readings, Wáng’s 1980 Shījīng readings, etc.) make the data comparable, both for humans and machines make the data freely available on the web and on Zenodo 12 / 35
Introduction Data Data: The CHIP Initiative (with N. W. Hill) try to digitize rhyme judgments of different authors (like Baxter’s 1992 Shījīng readings, Wáng’s 1980 Shījīng readings, etc.) make the data comparable, both for humans and machines make the data freely available on the web and on Zenodo Current state-of-the-art: http://github.com/digling/rhymes. 12 / 35
Rhyme Networks Modeling Example: The Shījīng Rhyme Browser interactive web-based application displays Shījīng rhymes in digitized form with rhyme annotations following Baxter (1992) and rhyme readings following Baxter and Sagart (2014) and Pān (2000, as provided in the Thesaurus Linguae Sericae). offers a quick and transparent way to inspect Baxter’s rhyme annotations, as well as a quick way to search through the Shījīng for rhyme patterns and brief glosses. URL: http://digling.org/shijing 16 / 35
Rhyme Networks Testing Testing: Vowel Purity (List et al. in press) Ho (2016) claims that the principle of vowel purity was important in Old Chinese rhyming: poets would try to avoid rhyming words with different consonants, 17 / 35
Rhyme Networks Testing Testing: Vowel Purity (List et al. in press) Ho (2016) claims that the principle of vowel purity was important in Old Chinese rhyming: poets would try to avoid rhyming words with different consonants, while differences in the codas were more often tolerated 17 / 35
Rhyme Networks Testing Testing: Vowel Purity (List et al. in press) Ho (2016) claims that the principle of vowel purity was important in Old Chinese rhyming: poets would try to avoid rhyming words with different consonants, while differences in the codas were more often tolerated reconstruction systems which contradict this principle, may therefore be externally criticized as neglecting the principle of vowel purity 17 / 35
Rhyme Networks Testing Testing: Vowel Purity (List et al. in press) Ho (2016) claims that the principle of vowel purity was important in Old Chinese rhyming: poets would try to avoid rhyming words with different consonants, while differences in the codas were more often tolerated reconstruction systems which contradict this principle, may therefore be externally criticized as neglecting the principle of vowel purity On the other hand, we can compare different reconstruction regarding the degree of purity of their vowels compared to the rhyme data in the Shījīng. 17 / 35
Rhyme Networks Testing Testing: Comparative Data Reconstruction System No. Rhymes Density a ɑ æ e ə o ɔ u ʊ ɯ i Karlgren (1957) 1830 0.0031 0.0026 x x x x x x x x x x x Li 李方桂 (1971) 1830 0.0031 0.0026 x x x x Wáng 王力 (1980) 1830 0.0031 0.0026 x x x x x Zhèngzhāng 鄭張尚芳 (2003) 1830 0.0031 0.0030 x x x x x x Starostin (1989) 1358 0.0035 0.0026 x x x x x x Pān 潘悟雲 (2000) 1830 0.0031 0.0026 x x x x x Baxter and Sagart (2014) 1431 0.0038 0.0033 x x x x x x Schuessler (2007) 1224 0.0041 0.0035 x x x x x x 18 / 35
Rhyme Networks Testing Vowel Purity: Methods Assortativity tests whether nodes sharing connections in a graph are also similar regarding other characteristics (New- man 2003). In social network analyses it can, for exam- ple, be used to test whether observed patterns in a network, like friendship, come along with properties of the individuals, such as language or gender (ibid.). Assortativity can be mea- sured by calculating the assortativity coefficient of a network in which all nodes have a given attribute. 19 / 35
Rhyme Networks Testing Testing Vowel Purity: Methods A 1 2 3 4 5 6 B 1 2 3 4 5 6 1 x x 1 x x 2 x x 2 x x 3 x x x 3 x x x x 4 x x x 4 x x x 5 x x 5 x x x 6 x x 6 x x x x 20 / 35
Rhyme Networks Testing Testing Vowel Purity: Methods A red blue red + blue B red blue red + blue red 6/14 = 0.43 1/14/ = 0.07 7/14 = 0.5 red 6/18 = 0.33 4/18/ = 0.22 10/18 = 0.55 blue 1/14 = 0.07 6/14 = 0.43 7/14 = 0.5 blue 4/18 = 0.33 4/18 = 0.22 8/18 = 0.44 red + blue 7/14 = 0.5 7/14 = 0.5 14/14 = 1.0 red + blue 10/18 = 0.55 8/18 = 0.44 18/18 = 1.0 20 / 35
Rhyme Networks Testing Testing: Results Provided the principle of vowel purity was really dominant during time of the creation of the Shījīng, our results indicate that reconstruction systems with six vowels outperform those with less or more vowels. Given that we do not know to which degree vowel purity was important in Old Chinese rhyming, this does not allow us to prove or disprove any of the reconstruction systems. Further research on rhyming practice and pragmatics are needed. 22 / 35
Rhyme Networks Testing Testing: Appendix One obvious criticism in the vowel purity paper remains. We tested on Baxter’s (1992) rhyme judgments, which could have easily influenced the results in the favor of the Baxter-Sagart system and of six-vowel systems in general! 23 / 35
Rhyme Networks Testing Testing: Appendix One obvious criticism in the vowel purity paper remains. We tested on Baxter’s (1992) rhyme judgments, which could have easily influenced the results in the favor of the Baxter-Sagart system and of six-vowel systems in general! We defended this with availability of data: when making the initial studies, no digital version of alternative rhyme judgments were available. 23 / 35
Rhyme Networks Testing Testing: Appendix One obvious criticism in the vowel purity paper remains. We tested on Baxter’s (1992) rhyme judgments, which could have easily influenced the results in the favor of the Baxter-Sagart system and of six-vowel systems in general! We defended this with availability of data: when making the initial studies, no digital version of alternative rhyme judgments were available. This is different now with the CHIP initiative, which just managed to digitize Wáng’s (1980) rhyme judgments along with the reconstructions. 23 / 35
Rhyme Networks Testing Testing: Appendix We can now compare the difference in rhyme judgments in Baxter (1992) and Wáng (1980). 1I excluded those with multi-morpheme rhymes. 24 / 35
Rhyme Networks Testing Testing: Appendix We can now compare the difference in rhyme judgments in Baxter (1992) and Wáng (1980). A simple measure is to compare, how many stanzas differ. From 1014 common stanzas1, 131 are different between Wáng and Baxter (12.9%). 1I excluded those with multi-morpheme rhymes. 24 / 35
Rhyme Networks Testing Testing: Appendix We can now compare the difference in rhyme judgments in Baxter (1992) and Wáng (1980). A simple measure is to compare, how many stanzas differ. From 1014 common stanzas1, 131 are different between Wáng and Baxter (12.9%). A far more useful measure is to compare how much different stanzas differ. Comparing rhyme judgments with a cluster task, B-Cubed scores are the perfect measure (Amigo et al. 2009). Applying B-Cubed scores to compare the rhyme judgments, we find 96.8% of similarity between Baxter’s and Wáng’s rhyme judgments. 1I excluded those with multi-morpheme rhymes. 24 / 35
Rhyme Networks Testing Testing: Appendix We can now compare the difference in rhyme judgments in Baxter (1992) and Wáng (1980). A simple measure is to compare, how many stanzas differ. From 1014 common stanzas1, 131 are different between Wáng and Baxter (12.9%). A far more useful measure is to compare how much different stanzas differ. Comparing rhyme judgments with a cluster task, B-Cubed scores are the perfect measure (Amigo et al. 2009). Applying B-Cubed scores to compare the rhyme judgments, we find 96.8% of similarity between Baxter’s and Wáng’s rhyme judgments. A new rhyme browser has now been created which contrasts rhymes by Wáng (1980) and Baxter (1992) and is available from http://digling.org/shijing/wangli/. 1I excluded those with multi-morpheme rhymes. 24 / 35
More on Networks Fǎnqiè Networks Fǎnqiè Networks It has long since been known that fǎnqiè 反切 readings can also be analyzed by exploiting their network characteristics (see, e.g., Gěng Zhènshēng 耿振生 2004 on the fǎnqiè xìliánfǎ 反切系聯法). 26 / 35
More on Networks Fǎnqiè Networks Fǎnqiè Networks It has long since been known that fǎnqiè 反切 readings can also be analyzed by exploiting their network characteristics (see, e.g., Gěng Zhènshēng 耿振生 2004 on the fǎnqiè xìliánfǎ 反切系聯法). But with modern network approaches, we can handle the data more consistently and transparently. 26 / 35
More on Networks Fǎnqiè Networks Fǎnqiè Networks It has long since been known that fǎnqiè 反切 readings can also be analyzed by exploiting their network characteristics (see, e.g., Gěng Zhènshēng 耿振生 2004 on the fǎnqiè xìliánfǎ 反切系聯法). But with modern network approaches, we can handle the data more consistently and transparently. By extracting, for example, all fǎnqiè shàngzì 反切上字 from the Guǎngyùn 廣韻, we can create networks of fǎnqiè connections. 26 / 35
More on Networks Fǎnqiè Networks Fǎnqiè Networks It has long since been known that fǎnqiè 反切 readings can also be analyzed by exploiting their network characteristics (see, e.g., Gěng Zhènshēng 耿振生 2004 on the fǎnqiè xìliánfǎ 反切系聯法). But with modern network approaches, we can handle the data more consistently and transparently. By extracting, for example, all fǎnqiè shàngzì 反切上字 from the Guǎngyùn 廣韻, we can create networks of fǎnqiè connections. These networks are ideal for teaching Chinese traditional phonology, but also for comparison if scholars have different opinions. 26 / 35
More on Networks Fǎnqiè Networks Fǎnqiè Networks Fǎnqiè networks are still underexplored, both with respect to traditional scholarship on Chinese historical phonology and with respect to the way they are best handled, and poten- tial differences across Chinese rhyme books or other sources containing fǎnqiè readings from different epochs or authors. However, it seems promising to further exploit and test the ap- proaches, as they may drastically increase the transparency of current approaches. 28 / 35
More on Networks Xiéshēng Networks Xiéshēng Networks Since Duàn Yùcái 段玉裁 detected the strong correlation between the phonetic part of xíngshēng 形聲 characters, we know that the Chinese system basically reflects a network structure, since a large part of the characters can be decomposed into subparts which reflect other characters or recur across different characters. 29 / 35
More on Networks Xiéshēng Networks Xiéshēng Networks Since Duàn Yùcái 段玉裁 detected the strong correlation between the phonetic part of xíngshēng 形聲 characters, we know that the Chinese system basically reflects a network structure, since a large part of the characters can be decomposed into subparts which reflect other characters or recur across different characters. Not often really taken into consideration is the historical aspect of these connections. Not all phonetic units of xíngshēng characters were formed at the same time, and the characters reflect a complex evolution of character formation at different steps. 29 / 35
More on Networks Xiéshēng Networks Xiéshēng Networks Since Duàn Yùcái 段玉裁 detected the strong correlation between the phonetic part of xíngshēng 形聲 characters, we know that the Chinese system basically reflects a network structure, since a large part of the characters can be decomposed into subparts which reflect other characters or recur across different characters. Not often really taken into consideration is the historical aspect of these connections. Not all phonetic units of xíngshēng characters were formed at the same time, and the characters reflect a complex evolution of character formation at different steps. Instead of listing xíngshēng series in form of lists of characters and common component, we should create explicit networks, as they are much more transparent to display where scholars disagree in their analyses, but also which characters are immediately composed of other characters. 29 / 35
More on Networks Xiéshēng Networks Xiéshēng Networks Displaying character structures, especially aspects of char- acter formation, with help of directed networks could greatly benefit not only scientific exchange among scholars, who would be encouraged to present their judgments more trans- parently, but also other aspects of Chinese writing, such as, e.g., pedagogical aspects of teaching the structure of the writ- ing system to beginners, or information-theoretic aspects. 31 / 35
More on Networks Dynamic Networks Dynamic Networks Not only the data we model in networks can be enhanced, also our methods to analyze the networks need to be further improved. As an example, consider dynamic networks, which would analyze and model network changes in time. By im- proving on these methods, we could, for example, compare fǎnqiè networks across different epochs, as well as rhyme networks from different authors, dialects, and styles. We could further try to induce fundamental hierarchies and rel- ative time frames from xiéshēng networks. 32 / 35
Outlook One of the major insights that I made during the last years in my research is that despite the great achievements schol- ars have made in historical linguistics, and especially in Chi- nese traditional phonology, we still lack clear-cut frameworks that help us to produce our data transparently. Historical lin- guistics is a data-driven discipline, but scholars tend to ignore this when presenting their incredible insights in an intranspar- ent form. Networks can help in two ways here: first, they are a transparent way of data-representation; and second, they provide an added value in those cases, where data becomes too large for scholars to be inspected by eye-balling only. 34 / 35
lingulist
rabernat
kamakiri1225
genomethica
joisino
ishiijunpei
s1ok69oo
ailaboocu
tiffany_04192
mkkon
ueniki
mlarrieu
62gerente
chriscoyier
samlambert
zakiwarfel
rasmusluckow
jakevdp
philhawksworth
tanoku
mclloyd
lara
geoffreycrofte
jacobian