Korpuslinguistik mit und f

1
Korpuslinguistik mit und für Computerlinguistik

• Martin Volk
• Universität Zürich
• Eurospider Information Technology AG

2
Sources for linguistic information

• Introspection (own usage and judgement)
• Usage and judgement by others
• Questioning (goal-driven)
• interview
• questionaire
• Observation ('involuntary' utterances)
• spoken utterances (? corpora)
• written utterances (? corpora)

3
What is a corpus?

• a text collection
• a representative text collection
• a representative and structured text collection
• a representative, structured and annotated text
  collection
• ...

4
Example

• Is 'ob' used as a preposition in German?
• Introspection
• Rothenburg ob der Tauber
• Dictionary (Wahrig. Deutsches Wörterbuch. 1996)
  Präp. mit Dativ veraltet ob dem Wasserfall
• Web Google 'ob dem'
• Sage Der Wilde Jäger ob dem Neuenburgersee
• Corpus

5
Corpus Examples

• CZ94 ... fiel schier vom Stuhl ob der Äusserung
  eines Ozeanologen ...
• CZ94 Bei manchem Ölgiganten kam ob der
  Ergebnisse gar Euphorie auf.
• CZ94 ... rieben sich vergnügt die Hände ob des
  zu erwartenden Schlagabtauschs.
• ob is a preposition with genitive!!
• in CZ corpus 'ob' is tagged as preposition 21
  times (obviously some incorrect)

6
History of Corpus Linguistics

• collections of text were widely used in the 19th
  century and in the first half of the 20th century
• language acquisition
• orthography (letter frequency)
• field linguistics
• ? American Structuralism (influential until 1960)

7
History of Corpus Linguistics

• Chomsky's criticism Speakers produce and
  understand infinitely many new sentences/words.
• therefore the new research goal is to describe
  the underlying language faculty of a speaker (
  universal grammar), competence rather than
  performance

8
History of Corpus Linguistics

• Chomsky's criticism every collection of texts is
  a collection of performance data and so many
  factors contribute to it that it cannot be used
  to model competence.
• A corpus is necessarily skewed. Some sentences
  won't occur because they are obvious, false or
  impolite.

9
History of Corpus Linguistics

• theoretical linguistics
• competence (what is grammatical?)
• introspection
• indefinitely many types, productivity
• grammatical vs. ungrammatical

• corpus linguistics
• performance (what is attested?)
• instances
• finite number of types
• degrees of grammaticality

10
Corpus research in Linguistics

• Lexicography (Dictionaries)
• Grammaticography (Reference grammars)
• Learner corpora Language acquisition
• Parallel corpora Translation

11
Construction of Corpora

• Written text is easier to obtain than spoken
  text. Some examples
• Newspapers
• Fiction (e.g. fairy tales)
• Technical Literature (e.g. manuals, medicine)
• Personal letters Email
• Advertising (incl. political propaganda)
• Belief and Thought (e.g. bible)

12
Corpora of spoken language

• Spontaneous spoken language
• recording of dialogues (e.g. telephone
  conversation)
• Prepared spoken language
• Public speeches (e.g. in parliament)
• Radio or TV news
• Spoken utterances must be transcribed for
  linguistic research.

13
Size of corpora

• Brown Corpus for English (1964, 1 Mio. words)
• LIMAS-Corpus for German (1970, 1 Mio. words)
• British National Corpus (1995, 100 Mio. words)
• Cosmas corpus (2002, gt 100 Mio. words)

14
Brown Corpus (1964)

• 500 texts
• out of 15 different text types
• with 2000 words each

15
British National Corpus

• 90 written English, 10 spoken English
• 3209 texts
• out of 10 different text types written and
• 6 text types spoken
• with lt 40'000 words each
• ? multi-purpose corpus

16
Other considerations

• Time frame of the corpus
• Native and non-native speakers
• Sociolinguistic variables
• Gender
• Age
• Education
• Dialect
• Social context and relationships

17
Types of corpora

• Raw texts
• Automatically annotated corpora
• Texts with Part-of-Speech tags
• Partially parsed texts
• Manually annotated corpora
• Treebank
• FrameNet

18
Types of Corpora

• Balanced Corpora vs. special corpora
• Spoken vs. written language
• Monolingual vs. Multilingual Corpora
• Parallel vs. comparable corpora

19
Corpora in Computational Linguistics
Corpora
annotation
Facts Rules Preferences
learning
20
My Motivation for Corpus Linguistics

• Attempt to build a parser for German
• But problems with ambiguities!!
• Therefore Learn attachment preferences from a
  corpus!

21
Corpora vs. Test suites

• A test suite
• is a collection of manually constructed and
  selected sentences.
• is used for testing computational grammars and
  parsers.
• reduces the amount of testing.
• leads to specific problems of the NLP system.

22
Basic problems in CL

• Knowledge is missing (too little information)
• e.g. unknown words
• Ambiguities (too much information)
• e.g. in syntax attachment preferences

23
Corpora in Computational Linguistics

• Widespread use of (manually) annotated material
  for measuring progress!
• Some examples from COLING 2002
• Treebanks to train and test probabilistic
  grammars
• Enriching treebanks with dependency information
• Automatic error detection in PoS-Tagged Corpora
• SENSEVAL data to train and test word sense
  disambiguation programs

24
Possible Student Tasks

1. Which German prepositions take a noun without a
   determiner? (e.g. pro, via)
2. When is mit used as an adverb? (e.g. )
3. What is the distribution of separable verb
   prefixes in German?
4. How often are relative clauses introduced with
   welche(r) ?
5. How often are present participle forms used in
   German?
6. What kind of foreign language material is in the
   corpus?

25
Possible Student Tasks

1. Create a small parallel corpus (e.g. with various
   versions of 'Alice in Wonderland' or National
   Geographic)
2. Create a small corpus of spoken language (e.g. by
   transcription of one issue of 'Big Brother').
3. Create a small treebank with the ANNOTATE tool.

26
What corpora do we have for German?

• Raw text
• ComputerZeitung 1993-97 (about 1.3 million words
  per year)
• ComputerZeitung iX
• Tages-Anzeiger 2000

27
Information in TagesAnzeiger

• Date
• Category (Sport, Politics, Culture, Economics
  etc.)
• Author
• Title vs. Text

28
What corpora do we have for German?

• Syntactically Annotated Text (Treebanks)
• NEGRA treebank (20'000 sentences)
• ComputerZeitung treebank (3'000 sentences)
• Text with manually corrected PoS tags
• 50'000 sentences from University speeches
• others

29
The goal

• If you can walk, you can dance.
• If you can talk, you can sing.
• If you can parse, you can understand.
• (Hans Uszkoreit, COLING 2002)

30
Acknowledgement

• Some slides were highly influenced by or even
  copied from Anke Lüdeling's course "Introduction
  to Corpus Linguistics" at http//www.cl-ki.uni-osn
  abrueck.de/aluedeli/Corpuslinguistik.html