Data in Linguistics

1
Data in Linguistics

• Linguistics -- by tradition -- is not an exact or
  empirical science. Modern linguists have
  attempted to transform linguistics step by step
  into a science.
• An exact science needs formalized models and
  provable methods for verifying (or more often
  falsifying) theories.
• Empirical science needs a methodology of how to
  obtain, process, evaluate data and how to exploit
  data for the verification (falsification) of
  theories.
• An exact empirical science needs to establish the
  correspondence between data and formal models.
  Therefore data need to be interpreted.
  Quantitative data require methods and tools for
  measurement.
• In linguistics, the quantitative branch of the
  discipline has been disconnected from the
  theoretical core of the field for many decades,
  since quantitative linguists could not measure
  phenomena that were in the focus of discussion.
  It was language technology that finally brought
  them together.
• Example Astronomy
• Photographs and spectral analyses of distant
  heavenly bodies are scientific data. However
  without their interpretation in relationship with
  the formal models, they are rather useless.

2
Data in Linguistics

• A well developed and established concept of
  linguistic data is still missing.
• No good theory of relationship between different
  types of data, e.g.,
• example sentences,
• online performance experiments,
• corpora,
• tree banks,
• test suites
• However,there has been progress in several areas,
  e.g.,
• evaluating acceptability judgements - methodology
  for subjective rating tests.
• annotation, interpretation of data
• methods for using quantitative data in language
  technology

3
Types of Linguistic Data 1

• Linguistically data are often classified into
  real and unreal data depending on their
  origin.
• However, this dichotomy does not fully cover the
  range of possible sources.
• naturally occurring data, e.g.,
• (balanced) reference corpora
• specialized corpora for specific subject domains
  or applications
• incidentally diccovered linguistic examples
• evoked or induced data , e.g.,
• dialogue-scenario data
• wizard-of-Oz data
• invented or solicited data, e.g.,
• sample sentences created by linguists
• acceptability judgements solicited by linguists
• test suites

4
Types of Linguistic Data 2

• The dichotomy real and unreal does not
  necessarily coincide with the property of
  naturalness.
• Linguistic examples are often considered
  unnatural. On the other hand, a large corpus
  may contain many sentences that are extremely
  unnatural.
• Naturalness does not solely depend on the origin
  of the data.

5
Concept of Linguistic Data

• If we view linguistics as an empirical science,
  pieces of linguistic knowledge have to be
  abstractions over linguistic data.
• These abstractions are parts of our theories
  about contents and structure of linguistic
  competence and about the processes, constraints
  and preferences that govern linguistic
  performance.
• Linguistic data are individual utterances, parts
  of utterances or collections of utterances in a
  certain human language (or several languages).
• The utterances may be represented in written or
  spoken form (or signed), i.e., as textual or
  acoustic signals.

6
Annotated Data

• Usually these collected utterances are annotated
  by additional information. If the annotation does
  not contain a partial linguistic interpretation
  of the utterances, the annotations may be
  considerd part of the data. Annotations that do
  not include linguistic interpretation are, e.g.
• judgements of native speakers on the
  acceptability or appropriateness of the
  utterance,
• information on speaker(s),
• information on hearer(s) or intended audience,
• information on the utterance situation (time,
  place, circumstances)
• information on the published source,
• typographic information,
• layout and document structure,
• textual transcriptions of spoken utterances,
• transcription of pauses.

7
Interpreted Data

• Annotations involving a partial linguistic
  interpretation are, e,g.
• part-of-speech tags,
• word sense information,
• morphosyntactic features of words,
• constituent structures for phrases or sentences,
  
• coreference markers,
• dependency structures,
• predicate-argument structures,
• reference identifications for term phrases,
• information structures within sentences,
• intonation contours,
• speech acts,
• discourse structures.

8
Parameters for Classification

• language Spanish, English, German
• sublanguage/register regional dialect,
  sociolect, vernacular, professional jargon,
  toddler speech
• text sort(s) newspaper articles, wire news,
  political speech, control commands
• subject domain stock rates, flight reservations,
  
• type of producers professional journalist,
  student, radiologist
• mode of production spoken, written, signed,
  morsed
• medium of production pencil, PC with MS Word,
  dictaphone
• conditions of production spontaneous, carefully
  composed, produced under time pressure
• transmission encoding raw ascii code, HTML,
  digitized phone signal, unicode
• medium of transmission telephone, WWW, CB radio
• storage encoding raw ASCII code, HTML, AIFF
• medium of storage DAT tape, CD ROM, hard disk
• mode of presentation spoken, written, signed
• medium of presentation newspaper, radio, book,
  tv show, theater performance,
• type of intended recipients newspaper reader,
  booking agent, theater audience
• number of intended recipients point-to-point,
  multicast, broadcast
• synchronicity of discourse synchronous dialogue,
  asynchronous
• direction one-way, two-way

9
Criteria for Usefulness

• In order to be useful, data have to be
  representative.
• representative of a certain linguistic
  phenomenon,
• representative of a certain text sort,
• representative of the expected input to some
  language technology application,
• representative of the expected output of some
  language technology application,
• representative of a certain speaker,
• etc.
• Can data be representative of an entire
  language?

10
Forschungsaufgaben

• Rohdaten sind heute einfach zu beschaffen.
• Die anspruchsvolle Aufgabe liegt in der
  linguistischen Interpretation.
• Forschungsaufgaben
• Entwurf der Annotationsschemata für die
  Beschreibungsebenen
• Entwurf von Austauschformaten und
  Übersetzungswerkzeugen
• Entwurf und Implementierung der Werkzeuge für die
  Korpusannotation
• Teilautomatisierung der Annotation
• Entwurf von Methoden und Werkzeugen für die
  Qualitätssicherung
• Entwurf von Werkzeugen für die Nutzung der Daten
  in der Forschung(Abruf, Auswertung, zusätzliche
  Dokumentation)
• Entwurf von Werkzeugen für die Nutzung der Daten
  für die Anwendungsentwicklung (Methoden und
  Werkzeuge für das Training)