Slides of the paper Standoff Annotation for the Ancient Greek and Latin Dependency Treebank by Giuseppe Celano at the 3rd Edition of the DATeCH2019 International Conference

1. STANDOFF ANNOTATION FOR THE ANCIENT
GREEK AND LATIN DEPENDENCY TREEBANK
DATECH2019, 9.5.2019
Giuseppe G. A. Celano

2. DFG PROJECT
1. Revise: correct the errors
2. Standardize: make the AGLDT standoff as PAULA XML (and convert into
UD)
1. standoff for multiple annotations and/or multiple interpretations of the
same token
2. standoff to overcome the problem of conflicting hierarchies
3. Expand: add new annotations
(https://git.informatik.uni-leipzig.de/celano/agldt1)

3. THE AGLT
▸ Ancient Greek texts: 557,922 tokens
▸ Latin texts: 79,697 tokens
▸ available in GitHub/GitLab:
▸ https://perseusdl.github.io/treebank_data/
▸ https://git.informatik.uni-leipzig.de/celano/agldt1

4. LABELED DIRECTED ACYCLIC GRAPHS

5. THE PERSEUS TREEBANK (LAST RELEASE, 2.1)
▸ 12 texts
composition date text token number
63 BC Cicero, In Catilinam 6,652
51 BC Caesar, De Bello Gallico 1556
post 44 BC Sallust, Bellum Catilinae 13191
ca 25 BC Prop. Elegiae 5297
29-19 BC Vergil, Aeneid 2839
ca 8 AD Ov., Metamorphoses 5209
14 AD Aug., Res Gestae 3035
15-50 AD Ph., Fabulae 6588
ca 100 AD? Petr., Satyricon 14177
ca 100-110 AD Tac. Historiae 3531
117-138 AD Suet., Vita Divi Augusti 8313
ca 400 AD Ger. Vulgata 9309

6. TREEBANK PIPELINE
start
Choose
a TEI/XML text
preliminary
automatic
annotation
tokenize it
rule-based
POS Tagger/Parser
manual
correction
end

7. THE PERSEUS TREEBANK: TEI XML TEXT

8. THE PERSEUS TREEBANK: INLINE ANNOTATION

9. INLINE ANNOTATION: ADVANTAGES
1. easy to add
2. easy to query
3. well supported by annotation tools

10. INLINE ANNOTATION: DISADVANTAGES
1. the tokenized text becomes the new base text
2. after text extraction from a TEI text, links to the original text is virtually lost
(e.g., amabam-que and content of some editorial markup)
3. it is unfeasible to connect such base texts to other annotation layers with
different tokenization schemes. For example:
‣ amabamque: one phonetic word
‣ amabam-que: two syntactic words
‣ am-a-ba-m-que: five morphemes
‣ verse vs. sentence

11. STANDOFF ANNOTATION
1. each annotation layer is attached separately to the original text
(i.e., the base text).
2. an annotation layer references the original text or another
annotaion layer which references the original text

12. STANDOFF ANNOTATION: PAULA XML
1. Open format based on the principles of LAF (ISO 24612:2012)
2. already employed in a number of historical language corpora
3. the base text is a bare xml text, which is virtually referenced only
via offsets

13. THE CASE STUDY: CAESAR’S DE BELLO CIVILI
1. the base text is a ‘complex’ TEI xml file’
‣ reference is made via XPath coinciding with CTS divisions
(https://git.informatik.uni-leipzig.de/celano/latinnlp/tree/master/case-study))

14. TOKENIZATION/WORD SEGMENTATION
▸ Latin: rule-based
▸ select the text to annotate from the TEI XML file
▸ identify abbreviations (word list + regular expressions)
▸ Cn. = Gnaeus
▸ list of not-to-tokenize words (e.g., Antigone, aeque)
▸ tokens ending with ne/que/ve
▸ list of to-tokenize words (e.g., nequis, nobiscum)

15. PAULA: TEI BASE TEXT

16. PAULA: TOKENIZATION

17. PAULA: SENTENCE SPLIT

18. CURRENT CHALLENGES
▸ extraction of text from TEI texts may require different scripts
▸ what is the ideal tokenization/word segmentation?
▸ annotation tools do not support standoff annotation
▸ lack of support for XPointer

19. THANK YOU FOR YOUR ATTENTION!