presentation of my research project held at the EIRI – CCH Conference on the Digitization in the Humanities at Keio University (Tokyo)

1. Structured Vs Unstructured:
Extracting Information From Scholarly Texts in
European Classical Studies
Matteo Romanello1
1 Centre for Computing in the Humanities
EIRI - CCH Symposium on the Digitization in the Humanities
Keio University - Tokyo 18th March 2010
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 1 / 26

2. Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 2 / 26

3. Introduction
Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 3 / 26

4. Introduction
The Project at a glance
Project started in October 2009;
Field of application: Digital Humanities, Classics (particularly
Greek literature);
co-supervision between the CCH and the CS department at King’s
-> application of Computational Linguistics method
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 4 / 26

5. Introduction
Focus
Scholarly Texts from the European Scholarly Tradition in Classical
Studies
Secondary sources, e.g. journal papers, as opposed to primary
sources, i.e. Ancient Texts
Sets of texts considered so far:
Princeton - Stanford Working Papers in Classics (PSWPC)
LEXIS online: classics journal available online under Open Access
policy
goal -> information extraction
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 5 / 26

6. Introduction
Goal
Devising an automatic system to improve semantic information
retrieval over a discipline-specific corpus of unstructured texts
focus on secondary sources
automatic -> scalable with huge amount of data
information retrieval -> the task of retrieving information
unstructured texts -> raw texts (e.g. .txt files) as opposed to the
structured/encoded XML
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 6 / 26

7. Motivations
Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 7 / 26

8. Motivations
The Million Book Library
archives.org, Google Books -> growth of
volume of information publicly available in
electronic format
longer “shelf-life” of books in
Classics/Humanities
need for effective tools to access
information for research purposes
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 8 / 26

9. Motivations
Information Extraction in Classics: challenges
lack of tools comparable to CiteseerX, GoPubMed, etc.
results of traditional search engines -> high recall but low precision
need to go beyond TOCs or string matching-based IR
still issues with encoding of Ancient Greek
no ad-hoc gold standards/training set
lack of tools specifically tailored to Classics resources
electronically available text does not mean electronic text
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 9 / 26

10. Methodology
Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 10 / 26

11. Methodology
Named Entities as Access Point to Information
mentions of entities matter for Classicists -> importance of print
indexes in Classics
Disambiguation, different spellings or translations of names
relating different expressions to the same entity
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 11 / 26

12. Methodology
Named Entities as Access Point to Information
Entities to be extracted:
1 Place Names (ancient and modern);
2 Relevant Person Names (mythological names, ancient authors,
modern scholars)
3 References to primary and secondary sources (canonical texts
and modern publications about them)
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 12 / 26

13. Methodology
Reuse of Structured Information
Reuse of structured data sources, e.g. thesauri, authority lists, etc.,
produced by scholars over the last two decades.
-> To train machine-learning based tools to mine unstructured texts.
Related work:
Research in the AI field -> Semantic Integration
Use of Wikipedia/DBpedia in NLP
Related projects: EROCS by IBM
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 13 / 26

14. Work Phases
Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 14 / 26

15. Work Phases
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 15 / 26

16. Work Phases
Corpus building
Getting materials
Crawling online archives
Extracting the text from collected documents
Tools for text extraction from PDF -> open issues with Ancient
Greek encoding
re-OCR documents even the native digital ones
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 16 / 26

17. Work Phases
Corpus Building II
Corpora
open access, multilingual
Princeton/Stanford Working Papers in Classics (PSWPC)
Lexis online
470 articles in 2 corpora
OCR
Finereader
Ocropus (layout analysis)
text extracted from PDFs (tools like pdftotext etc.)
Alignment of multiple OCR outputs
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 17 / 26

18. Work Phases
Building the Knowledge Base (KB)
Goal: integrate different data sources into a single KB
Why?
Information about the same entities spread over several data
sources
Data sources might use different output formats (raw text, DBs,
HTML, XML etc.)
partial overlappings but no interoperability
How?
Use of high level ontologies to map records related to the same
entity
Result: KB containing semantic data
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 18 / 26

19. Work Phases
Building the Knowledge Base (KB) II
Ontologies -> in CS a formalism to model data
Integrating data sources:
import each datasource
map it to high level ontologies (e.g., CIDOC-CRM)
find overlappings between datasources -> alignign the records
The obtained knowledge base will be used as support for all the text
processing tasks
Implementation of the KB: RDF triple store with a SPARQL interface
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 19 / 26

20. Work Phases
Corpus Processing
1 sentence identification
2 entities extraction (named entities recognition + disambiguation)
KB implied to build up an entity context
3 canonical references extraction
KB provides training data
4 modern bibliographic references extraction
KB provides list of journals/name places/authors to improve the
perfomances of the tool
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 20 / 26

21. Work Phases
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 21 / 26

22. Work Phases
Canonical References Extraction
1 citations used specifically for primary sources (i.e. works of ancient
authors)
2 essential entry point to information: refer to the research object, i.e.
ancient texts
3 logical instead of physical citation scheme (e.g., chapter/paragr vs.
page)
4 variation -> time, style, language (regexp insufficient!)
Example
Hom. Il. XII 1
Aesch. ’Sept.’ 565-67, 628-30; Ar. ’Arch.’ 803
Hes. fr. 321 M.-W.
Callimaco, ’ep.’ 28 Pf., 5-6
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 22 / 26

23. Expected Results
Overview
1 Introduction
2 Motivations and Background
3 Methodology
4 Work Phases
5 Expected Results
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 23 / 26

24. Expected Results
Results
Provide automatically multiple meaningful entry points to
information
Enrich the corpus with links to resources (particularly primary
sources)
Improve the user access to the corpus
Demonstrate the scalability of the approach
Tools/Resources
Knowledge Base for Classics
Articles with improved text quality
(improved) corpora to be released
single tools for information extraction (e.g. CREX Canonical
References EXtractor)
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 24 / 26

25. Expected Results
Possible Applications
Solution to problems peculiar of Classics might help to improve
the performances of existing tools/algorithms
Collections of secondary sources as corpora:
citation patterns
citation and co-citation networks
trends in the Classics citation practice
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 25 / 26

26. Expected Results
Thanks for your attention!
matteo.romanello@kcl.ac.uk
http://uk.linkedin.com/in/matteoromanello
Romanello (CCH) Extracting Information From Scholarly Texts EIRI - CCH Symposium 26 / 26