Outputs

Tibetan Modern U-chen Print 0.1

Posted on by Xaver in Corpora, models

Tibetan Modern U-chen Print 0.1 (TMUP 0.1) is the first Transkribus HTR model for printed Tibetan language publications in Uchen (དབུ་ཅན་ dbu can) script. It has been trained on texts that were published in the PRC between the 1950s and 1980s. The model was trained on 522 pages in 20 documents. The training set consists of 470 pages; the validation set consists of 52 (10%) automatically selected pages. No base model was used. The model was developed by Franz Xaver Erhard (Leipzig University) and Xiaoying 笑影 (Leipzig University) for the Divergent Discourses project (DFG/AHRC).

The model is publicly available within the Transkribus environment. You can view and test the model at

https://readcoop.eu/model/tibetan-modern-u-chen-print/

Details on the model and the Ground Truth of the training set can be viewed on the Transkribus-Site of the Divergent Discourses project:

https://app.transkribus.org/sites/uchan

The training set to the model – consisting of the image files (jpg) and the corresponding Transkribus pageXML files – is available for download from:

Erhard, Franz Xaver, Xiaoying 笑影, Barnett, Robert, Hill, Nathan W., 2024. Tibetan Modern U-chen Print (TMUP) 0.1: Training Data for a Transkribus HTR Model for Modern Tibetan Printed Texts. https://doi.org/10.48796/20240313-0000352

Tibetan_tokenizers: botok_tokenizer.py

Posted on by Xaver in Code

botok_tokenizer.py is a tokenization or word segmentation utility for Tibetan based on BoTok, a tokenizer developed by OpenPecha. It allows you to point to a whole folder or directory or to a single .txt file. It selects just the tokenizer element of BoTok rather than BoTok’s POS tagger element, which we have not included in this utility.

Engels, J., Barnett, R., Erhard, F., & Hill, N. (2024). Tibetan_tokenizers: botok_tokenizer.py (v1.1). Zenodo. https://doi.org/10.5281/zenodo.10810709

View code on Divergent Discourses Github

botok_tokenizer.py was developed by James Engels of SOAS University of London for the Divergent Discourses project. The project is a joint study involving SOAS University of London and Leipzig University, funded by the AHRC in the UK and the DFG in Germany.  Please acknowledge the project in any use of these materials. Copyright for the project resides with the two universities.

Transkribus_utils: Paragraph Extractor: A tool to extract text from Transkribus pageXML

Posted on by Xaver in Code

Transkribus transcribes the text on a given page line by line and doesn’t discriminate between different forms of formatting such as headings, marginalia or footnotes. To meaningfully structure and separate the transcribed text into smaller units is crucial for text and corpus analysis, e.g., with the Leipzig Corpus Miner (iLCM). To retrieve structured plain text from Transkribus pageXML, further processing is neccessary.

This repository holds utilities for parsing and extracting useful data from Transkribus PageXML outputs, such as a utility for identifying text regions (Paragraph Extractor), and a utility to reconcile Trankribus output metadata with the equivalent data in relevant library catalogues (coming shortly).

Engels, J., Robert Barnett, Erhard, F. X., & Hill, N. (2024). Transkribus_utils: Paragraph Extractor (v1_Paragraph_Extractor). Zenodo. https://doi.org/10.5281/zenodo.10810509

View code on Diverge Github

Paragraph Extractor is a utility that accepts Transkribus PageXML as input and then interprets the text regions on each page/image (such as headers, titles, blocks of text, etc.), which we term “paragraphs”. It then returns the raw text of each text region (paragraph) along with its metadata. Note that it reads PageXML, not AltoXML.

Paragraph Extractor was developed by James Engels of SOAS University of London for the Divergent Discourses project.

TibNorm: Script to Normalise Tibetan Text

Posted on by Xaver in Code

TibNorm is a utility for producing normalised versions of Tibetan texts to make them easier for contemporary users to search and read, in line with current Tibetan writing conventions.

Kyogoku, Yuki, Robbie Barnett, & Franz Xaver Erhard. (2024). TibNorm – Normaliser for Tibetan (Version v1). Zenodo. https://doi.org/10.5281/zenodo.10806456

See code on Diverge github

As part of the normalisation process, TibNorm:

  • changes Tibetan numbers into Arabic numerals
  • changes Tibetan brackets and quotation marks into the standard western equivalents
  • removes a ། if found after a ཀ, ག or ཤ, with or without a vowel – adds a ་ between ང and །
  • reduces two or more ་ to a single one
  • changes ཌ་ or ཊ་ to གས་ unless preceded by a white space, tab, or new line
  • changes non-standard “illegal” stacks into standard ones
  • deletes a ། if found at the beginning of a line

TibNorm also expands abbreviations so that they are shown in their full form. For abbreviations in classical Tibetan, TibNorm draws from the list of over 6,000 classical Tibetan abbreviations compiled by Bruno Lainé of the Tibetan Manuscript Project Vienna (TMPV) as part of the project’s Resources for Kanjur and Tanjur Studies. In TibNorm, the user can manually change the flag in the abbreviations table to exclude any abbreviation that they don’t want to expand.

TibNorm was developed for the Divergent Discourses project by Yuki Kyogoku of Leipzig University.

Tibetan language support in iLCM

Posted on by Xaver in Publications

The following dataset consists of a Tibetan language model for SpaCy and a list of Tibetan stopwords to enable Tibetan language support in the integrated Leipzig Corpus Miner (iLCM); it includes instructions to upload the model into the iLCM framework.

Engels, James, Erhard, Franz Xaver, Barnett, Robert, & Hill, Nathan W. (2023). Tibetan for Spacy 1.1 (1.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10120779.

A significant obstacle to using major state-of-the-art NLP applications in Tibetan studies is the lack of support for the Tibetan language. Since the DIVERGE project aims to analyse several thousand pages of Tibetan newspapers, it depends on applications such as the integrated Leipzig Corpus Miner (iLCM).

The iLCM is an interface-based text and corpus mining software package capable of a range of NLP tasks, such as frequency analysis and topic modelling. The iLCM is a wrapper on functions accessible through SpaCy but avoids the scripting requirement of a boilerplate spacy model. It requires a small language model built from the SpaCy NLP package as input. SpaCy has native support for several high- and high-intermediate resource languages and is an industry-standard software package for small-scale English and other European-language NLP integration in a variety of research and corporate environments. No major small language model NLP package available today has native support for Tibetan at any step of the pipeline.

Only little training data for a language model is available for Tibetan as a low-resource language. Nevertheless, to get started, we have developed a preliminary Tibetan language model for SpaCy with limited available training data in CoNLL-U, preprocessing it with BoTok.

The iLCM does not have a native Tibetan model and does not natively support every language supported by SpaCy but is packaged with a few common European languages (English, French, German, Italian, Spanish, and Portuguese). A new model must be separately uploaded with a separate document including an explicit list of stopwords (generally semantically vacuous function words that are not useful for tasks like topic modelling).