Dirk Goldhahn (University of Leipzig, NLP group) was the only speaker presenting a linguistic dataset from the academic field. He introduced the the Leipzig Corpora Collection. The dataset comprises corpus-based full form monolingual dictionaries for more than 220 languages which comes with a variety of meta-data, e.g. word frequencies, POS tagging and co-occurrences. Furthermore, the corpora are enriched with statistical annotations such as POS, topics, word and co-occurrence frequencies. At the moment the NLP group is working on a conversion of their data into a linked data format. At the same time integration work of external sourced still needs to be done.

1. Processing Extensive Text Data
at the Leipzig Corpora Collection
Dirk Goldhahn
MLODE 2014, Leipzig
Natural Language Processing Group
Institute of Computer Science
University of Leipzig

2. Leipzig Corpora Collection
2
● Project that offers
● Corpus-based monolingual full form dictionaries
● Started as project „Deutscher Wortschatz“
● Accessible at http://corpora.uni-leipzig.de
● Web interface
● Web services
● Integration into CLARIN
● Export as rdf (Sebastian Hellmann)

3. Available Resources (Online)
3
● Corpora in >220 languages
● Text material
● Word frequencies
● Word co-occurrences
● Semantic maps (strongest word co-occurrences)
● Topics
● Subject areas (partially)
● Morphological information (partially)
● POS-tagged sentences (partially)
● Further semantic and grammatical information

4. Example
4

5. Example
5

6. Number of Languages over Time
6

7. World Map Interface
7

8. Data Stock
●Number of languages: >300, >1000 including religious texts
●Number of sentences (overall): ~4 billion
●Biggest language: English (1.1 billion)
8

9. Text Acquisition Methods
9
●Generic web crawler (Heritrix)
● Whole TLDs
● News (abyzNewsLinks)
●RSS feeds (daily)
→ http://wortschatz.uni-leipzig.de/wort-des-tages/
●Distributed text crawling (FindLinks)
●Bootstrapping web corpora
●Text dumps (Wikipedia, Watchtower)
●Parallel corpora (Bible)

10. Corpus Creation
10
●Processing steps:
● Stripping
● Language Separation (about 500 languages)
● Sentence Segmentation
● Cleaning
● Sentence Scrambling (copyright reasons)
● Tokenization
● Database Creation
● Statistical Evaluation

11. Statistical Evaluation
11
●Enrichment of corpora with statistical annotations
● Word frequencies
● Co-occurrence frequencies and significance (based on left or right
neighbours or whole sentences)
● Topics, POS
●Creation of further corpora statistics
● >200 statistical properties based on letter, word, sentence, sources level
etc.
● e.g.: For automatic quality assurance using typical distributions
Hindi newspaper 2010 %
0 50 100 150 200 250
1
0,8
0,6
0,4
0,2
0
0 50 100 150 200 250
6
5
4
3
2
1
0
Sundanese Wikipedia 2007
%

12. LCC as Linked Data
12
●LD starts with data and metadata
● Data + metadata for hundreds of languages
● All data processed and stored identically,
including metadata
● Still at the beginning of turning into LD

13. LCC as Linked Data
13
●Monolingual Metadata:
● Frequencies, word co-occurrences, NER, POS-tags, topics
● Integration of external sources (DBPedia, FreeBase, Wordnet)
● Still problems to solve (linking, disambiguation, ...)
→ New Web Interface
●Multilingual Metadata:
● Crosslingual word co-occurrences (parallel text)
● External sources (DBPedia...)

14. 14
●Data available via:
● Web interface
(http://corpora.informatik.uni-leipzig.de)
● Download
(http://corpora.informatik.uni-leipzig.de/download.html)
● Webservices
(http://wortschatz.uni-leipzig.de/Webservices/)
Thank you!