Effective Identification of Citations in the Kanseki Repository

poster / demo / art installation
  1. 1. Christian Wittern

    Kyoto University

Work text
This plain text was ingested for the purpose of full-text search, not to preserve original formatting or readability. For the most complete copy, refer to the original conference program.

The Kanseki Repository is a large repository of premodern Chinese texts. Currently it holds more than 9000 texts, covering all periods of Chinese history from early antiquity to the beginning of the 20th century. The repository is organized into 6 top-level categories and offers full-text search across all textual variants.

Since its opening to the public in March 2016, a frequent request from users was to be able to find texts related to a certain text, especially to investigate and evaluate textual dependencies. This poster reports on some experiments to find an effective way to respond to this requests.

What is needed to solve this problem is an efficient way to identify text passages that are derived from other, earlier texts (based on the assumption that the texts in question can in fact be reliably dated). Such passages will be called citations here, although the usage here is not limited to true citations, but also includes quotations from memory, paraphrases and allusions - cases where the reference does not follow the exact wording of the source. As an additional complication, we need to take into account the possibility that the received text differs from the text available to the author of the text that contains the reference.

Related work has examined plagiarism detection (Gipp and Meuschke, 2011 and Schultz, 1999), but the approach taken here makes direct use of some of the unique features of the repository and the index built for it and seems thus to be more efficient than general purpose algorithms. Admittedly, this has not been verified empirically and thus may be reasonably rejected as not relevant. However, the purpose of this presentation is not to compare algorithms and their efficiency when applied to the material here, but rather to collect some low-hanging fruit that became available due to the way the full-text index is constructed.

Identification of Citations

Since a complete index has already been built for the full-text search, all experiments make use of this index (Mandoku 2016). The index is constructed by moving an n-gram window over the text and saving entries at appropriate locations. The resulting raw index is then read by a grep-like program to generate the display. The search display is designed to show the keyword in context (KWIC) so some characters are needed in front of the match character; these are appended after a comma character in the index entry. We built the index with a window between 10 and 25, since larger indexes considerable increase the required space and smaller builds will have too little information in the KWIC display. The index also contains information to identify the text, the location of the index excerpt and some information about the context of the match. Figure 1 shows a typical example of such an index for a 21-gram window.

Æ M+-it £ *£, I % #4+

« && % * a , w# m+


« MX« 4Æ. ür A- * < , A if «

«&KR0015 ; A&KR0003 ; W, a

WMitWO#®, £& B

AA Aits,s' , AAAfezr

^,7} it #A® <7} a A S- , - A BJ#-

A AiS & # A A S' < it A, A « Si B

t At

W , KRla0003_001: 6a

U:%%- 1% (123,0) (Fundamer

KRla0002_009: 7b : 2:16:2 KRla0002_009: 7b : 3:6:2 KRla0002 011: la : 2:7:0 KRla0002 011: la : 4:6:2

KRla0003 000:la:3:2:0 KRla0003 000:la:3:6:0 KRla0003 000:2a:4:12:0

KRla0003 000:2a:4:12:0 5BCK KRla0003 000:1a:1:9:0 WYG KRla0003 000 : la : 1:30:2 WYG KRla0003 000:la:1:161:10 WYG

KRla0003 000:1b:4:40:24 WYG KRla0003 001:lb:2:4:3 n KRla0003 001:3a:5:16:8 KRla0003 001:3a:6:16:8 KRla0003 001:3a:7:1:8 KRla0003 001:8b : 9:20:23 KRla0003_001:9b:l:18:25 :8:3:17 n

KRla0003 001:11a:8:4:29_

ital +1) Mon Oct 31 5:13PM 0.26

Figure 1. Effective Identification of Citations in the Kanseki Repository


To find citations in the indexed files, a window of the same size as the index window is moved over the text passage under investigation. A search is initiated for a string of n characters, starting at the first index position. In the example in Figure 1 this would be starting at position 6, since there are 5 characters after the ","; these characters are preceding the indexed characters in the text and have therefore been moved to the back. A query expansion is used for this search, in order to catch character variants in this initial selection of index lines. A large value of n will increase the probability that citations are not found due to slight positional variations in the text, while a small value of n, such as 1, will select many lines that are not relevant and will thus increase the processing times. Experiments have shown that a value of 2 or 3 for n gives a good optimum for precision and recall. Positional values are also registered to better demarcate the citation boundaries.

The selected lines have to be post-processed to restore the original sequence as found in the text. The line will then be compared to the window of the text passage, with scores given for each match; high scores are taken to be a citation and are marked for further processing. The best results so far have been achieved for a cumulated score of n-gram matches for values of n from 1 to 3, but additional experiments are planned. Conclusive results will be reported in the poster presentation.

Additional expected results
With the method introduced here, it becomes feasible to investigate potential citations for whole texts. We plan to build a heat-map of a text with passages that have been cited coloured according to their frequency. This will enable new ways of exploring the intertextuality of texts and will provide new evidence for the history of ideas and flow of intellectual debt in the history of Chinese thought. For the presentation of the poster, we show a preliminary investigation of some key texts of Chinese philosophy as a proof of concept.

We also hope to identify a set of key phrases and look at their usages over time, and in different schools of thought, to see what kind of trends can be seen there.

Wittern, C. (2014). "Kanripo and Mandoku: Tools for git-based distributed repositories for premodern Chinese texts", in Digital Humanities 2014 Book of Abstracts, 2014, p. 408-409.

Wittern, C. (2016). Special issue Kanseki Repository, CIEAS Research Report 2015, Kyoto 2016.

Gipp, B. and Meuschke, M. (2011). "Citation Pattern Matching Algorithms for Citation-based Plagiarism Detection: Greedy Citation Tiling, Citation Chunking and Longest Common Citation Sequence". In: Proceedings of the 11th ACM symposium on Document engineering (DocEng ’11), Mountain, View, CA, USA, 2011. doi: 10.1145/2034691.2034741.

Mandoku (2016). Mandoku project (Source code) https://github.com/mandoku/mandoku (accessed 2016-10-31).

Schultz, R.L. (1999) The search for quotation : verbal parallels in the prophets, Sheffield, Sheffield Academic Press, 1999.

If this content appears in violation of your intellectual property rights, or you see errors or omissions, please reach out to Scott B. Weingart to discuss removing or amending the materials.

Conference Info


ADHO - 2017

Hosted at McGill University, Université de Montréal

Montréal, Canada

Aug. 8, 2017 - Aug. 11, 2017

438 works by 962 authors indexed

Series: ADHO (12)

Organizers: ADHO