Stylometric Analysis of Chinese Buddhist texts - Do different Chinese translations of the Gaṇḍavyūha reflect stylistic features that are typical for their age?

May combine two claims.

In the PCA results for 27 to 24 particles (Figs. 9-12), G-Jin units cluster consistently and distinctly, and G-Tang-1 and G-Tang-2 are closer to each other than to G-Jin.

“The first five results in this series (Figs.9-12) are encouraging.G-Jin units cluster consistently and distinctively and G-Tang-1 and G-Tang-2 are closer to each other than to G-Jin.”

May combine two claims.

Experiment 3 suggested G-Jin's use of the 33 particles differed from the Tang versions, but most units stayed close, the clusters were unclear, and two outlier groups distorted the outcome.

“Although Experiment 3 appears to cluster G-Jin, giving an indication that its use of the 33 particles differs from the two Tang dynasty translations, the majority of units are still close together, the clustering is indistinct, and there are two sets of outliers that distort the result.”

May combine two claims.

The particles 爾, 故, and 時 left the outliers' influence unchanged and made clustering even less clear.

“Fig. 6 shows that the addition of 爾, 故, and 時 does not change the power of the outliers, if anything, the clustering has become more indistinct.”

May combine two claims.

In Figs. 16 and 17, G-Jin units cluster and move toward G-Qin.

“The interesting part is that in Fig. 16 and Fig. 17 the units of G-Jin are clustering and moving toward G-Qin”

• AssumptionRests on an assumption: The Luomaqiejing (G-Qin, T. 294) is a partial translation of the Gaṇḍavyūha and is traditionally ascribed to Shengjian.
• AssumptionRests on an assumption: G-Qin is nearly contemporaneous with G-Jin.

May combine two claims.

With 10 units per text, PCA placed 7 of 10 G-Jin units in the lower-left area; this division gave clearer clusters and was adopted for later steps.

“Only in Experiment 3 (Fig. 3) the PCA clusters the majority of G-Jin units (7 out of 10) in the same region of the plotted graph (lower left).Dividing the texts in 10 units each seems to yield clearer clusters and this number is used for the following steps.”

Classified as probably not a substantial claim.

The first two criteria-high frequency and high inter-textual variance-produced a set of 33 particles.

“The first two conditions have yielded the set of 33 particles, and Table 5 presents the set sorted for the third condition.”

May combine two claims.

In the internal-variance ranking, 者 is at the top, while 故, 為, and 諸 also show high internal variance.

“者, not surprisingly, tops the list, but other particles such as 故，為，or 諸 also display high internal variance.”

May combine two claims.

With 40 units per text, all units formed one cluster, indicating PCA could not handle 120 units; this division exceeded the threshold for distinct clustering and was unusable.

“The result of the first experiment (40 units per text) (Fig. 1) shows all units clustered together in a way that suggests that PCA cannot cope with the number of units.120 appears to be above the limit that allows the analysis to result in a distinct clustering and dividing the texts in 40 units each is therefore not going to work.”

May combine two claims.

Raising the list to 25 particles did not clarify the result; while most G-Jin units sat in the lower left, the two outlier groups still controlled the distribution.

“In Experiment 7 the number of particles is raised to 25 but the result does not become clearer.In Fig. 7, although most G-Jin units cluster clearly in the lower left corner, the two outlier sets still determine the distribution.”

May combine two claims.

At 20 units per text, G-Jin began to separate into its own cluster, though the separation was still unclear.

“In Experiment 2 (Fig. 2) each text is divided in 20 units and G-Jin seems to start forming an independent cluster but the clustering is still indistinct.”

May combine two claims.

Within the 10 units, 者 showed the greatest internal variance of any particle, occurring unusually often in certain units.

“of all particles zhe 者 has the highest internal variance among the 10 units, i.e. it appears disproportionally more often in some units.”

May combine two claims.

When the ten particles with the highest internal variance are excluded (Fig. 12), the two persistent outliers join the fold; they are not quite in the main G-Jin cluster but no longer distort the output.

“Excluding the 10 particles with the highest internal variance factor, as shown in Fig. 12, even the two persistent outliers come into the fold, and, while still not quite part of the main cluster of G-Jin, they are not distorting the output anymore.”

Probably not a substantial claim: may not pass S3 (load-bearing).

G-Qin is less than one quarter the size of G-Jin.

“G-Qin is nearly contemporaneous with G-Jin, however, it is less than 25% of its size.”

May combine two claims.

With 28 particles, G-Jin forms a clear separate cluster, and the units of G-Tang-1 and G-Tang-2 appear to split into two clusters.

“Based on the use of these 28 particles G-Jin clearly forms a discreet cluster, and even the units of G-Tang-1 and G-Tang-2 seem to divide in two different clusters.”

May combine two claims.

The 11th particle, 者, produced unexpected effects: unit distances shrank across translations, G-Jin still clustered visibly, and the two outlier groups grew stronger, worsening the analysis.

“In experiment 5 the addition of the 11 th particle '者' results in unexpected changes.In Fig. 5, Distances between units of all translations have decreased though the clustering of G-Jin is still discernible (lower left).It seems that the addition of 者 has strengthened the two sets of outliers and distorts rather than helps the analysis”

May combine two claims.

其 has a particularly high external variance value and high frequency.

“We have seen in Table 1 that this character has a particular high external variance value as well as high frequency”

May combine two claims.

Excluding 其 (Fig. 13) appears to reach a threshold, and the clustering becomes less distinct again.

“With the exclusion of 其 (Fig. 13), however, a threshold seems to be reached and the clustering becomes less distinct again.”

Classified as probably not a substantial claim.

The 23 particles are the high-frequency particles from Table 5, excluding the top ten that are used unevenly across units.

“These are the high-frequency particles listed in Table 5, without the top ten which are used unevenly across the units.”

Classified as probably not a substantial claim.

Removing the particles with the greatest internal variance from the list should lessen distortion.

“Taking the particles with the highest internal variance off the list should reduce distortion.”

Probably not a substantial claim: may not pass S1 (asserted), S4 (based).

A finding that G-Qin and G-Jin group together would suggest the method can detect chronological translation features.

“If our method would show that G-Qin and G-Jin are related, it would be an indication that it does distinguish features due to the time of translation.”

• AssumptionRests on an assumption: Particle usage is characteristic of the period of translation, so differences in particle frequencies reflect translation date rather than translator or content.

Classified as probably not a substantial claim. May combine two claims.

Segmenting Chinese into words is hard even for modern Chinese, and separating compounds from words is notoriously difficult.

“The difficulty to disambiguate phrase level compounds and words has been called”

• cites Norman 1988: 156Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Stylometry favors non-contextual items over items that draw human attention; style shows in function words, common verbs, and punctuation, not in rare literary nouns and adjectives.

“It has therefore long been held that non-contextual words or phrases are better features for stylometric analysis than words or phrases that human readers tend to focus on.”

• Also stated in: § １１ · plain

  “Style is encoded in the use of features such as grammatical particles, basic verbs or even punctuation rather than in the literary use of rare, but”

• cites Mosteller and Wallace 1984 sourceNot checked — no full text of the source
• cites Zhao and Zobel 2007Not checked — no full text of the source

Classified as probably not a substantial claim.

An optimal particle set consists of items that are fairly frequent in every text and vary most across texts.

“the best list of particles contains all those particles that are used relatively frequently across all texts, and at the same time have maximum variance between texts.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

There is no established way to choose the best unit count, so the authors had to test different numbers.

“As of yet we have no exact method to determine the optimal number of units and therefore have to experiment in order to see which amount of units yields the clearest results.”

Classified as probably not a substantial claim.

When unsupervised clustering matches a hypothesis, it counts as strong evidence, since no pre-labeled training categories shaped the grouping.

“If the results of the analysis turn out to confirm a hypothesis of how the texts would cluster these types of analysis amount to strong evidence for the hypothesis, because the clustering has not been influenced by pre-conceived categories in a set of training data.”

May combine two claims.

Chinese grammar relies on syntax and a small inventory of function words, and the employment of those words shifted markedly across periods.

“Grammatical information in Chinese is expressed through syntax and with the help of a limited number of grammatical particles (xuzi 虛字), the use of which changed considerably over time.”

• cites [7] Dobson, William A. C. H. (1962). Early Archaic Chinese – A Descriptive Grammar. Toronto: University of Toronto Press.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

T.295, another Gaṇḍavyūha text in the canon, is just a fragment and is unusable here.

“Another Gaṇḍavyūha text in the canon (T.295) is merely a fragment and cannot be used here.”

• cites T.295Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

The three full translations reflect different stages of the Sanskrit text, vary in length, but broadly match in content and structure.

“The three full translations represent different stages in the development of the Sanskrit text and differ in length, but agree in general as to their content and structure.”

• AssumptionRests on an assumption: The three translations are sufficiently comparable in content and structure for their particle usage to be compared stylistically.

May combine two claims.

Although frequent features often perform well, the structure of literary Chinese makes very frequent words a poor choice.

“This works well in many cases, but unfortunately the structure of literary Chinese, whether in its Buddhist or more”

• cites Burrows 1992 sourceNot checked — no full text of the source
• cites Hoover 2002 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based).

G-Qin is close in time to G-Jin.

“We will bring G-Qin, which is close in time to G-Jin, into play only later”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

PCA responds strongly to how units are scaled, and the count of units strongly affects result quality.

“PCA is sensitive to the scaling of the units and the number of units is critical for the quality of the results.”

Classified as probably not a substantial claim.

This study uses function words, which together with syntax carry most grammatical information in classical Chinese.

“Our approach here focuses on grammatical particles (xuzi 虛字), which, next to syntax, are the main carriers of grammatical information in classical Chinese.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In content, G-Jin and G-Tang-1 closely correspond, whereas G-Tang-2 derives from a much longer Sanskrit version.

“Content-wise, G-Jin and G-Tang-1 are very close, while G-Tang-2 was translated from a considerably expanded Sanskrit text (Gómez 1967: xxvi-xxviii).”

• cites Gómez 1967: xxvi-xxviiiNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: The three translations are sufficiently comparable in content and structure for their particle usage to be compared stylistically.

May combine two claims.

Stylometry can process long texts without sampling and evaluate vastly more features than a human reader can.

“Stylometric analysis can attempt to address longer texts, without the need to sample.Crucially, it is now possible to test a large amount of features, orders of magnitude beyond the reach of the unaided human reader.”

May combine two claims.

Buddhist Hybrid Chinese frequently uses multi-character compounds, particularly names and doctrinal terms, and particles inside them would bias the analysis.

“In Buddhist Hybrid Chinese, on the other hand, compounds of two or more characters are quite common, especially for names and doctrinal terms.The grammatical particles that we are interested in can also appear as part of such compounds which would distort the analysis.”

May combine two claims.

G-Jin belongs to T.278, a Buddhāvataṃsaka version produced by Buddhabhadra's team in Chang'an during 418-420 CE.

“As part of the Buddhāvataṃsaka (T.278) translated by Buddhabhadra 佛陀拔陀羅 et al. in Chang'an between 418 and 420 CE in the final years of the Eastern Jin dynasty”

• cites T.278Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

If Jin and Tang particle usage were not distinct, then no particle selection should yield a clear separation between the texts.

“If there is no clear distinction in the way particles are used in the Jin vs. the Tang dynasty one would expect that no set of particles would result in a clearly discernible difference between texts.”

• AssumptionRests on an assumption: Particle usage is characteristic of the period of translation, so differences in particle frequencies reflect translation date rather than translator or content.

May combine two claims.

The difference in particle use is due not only to the idiosyncrasy of the translations but also to the fact that they were produced at different times.

“This is an indication that the difference in particle use is not only due to the idiosyncrasy of the translations, but also to the fact that they were produced at different times.”

• AssumptionRests on an assumption: G-Qin is nearly contemporaneous with G-Jin.