Discourse processing

1
Discourse processing NLP

• Sofia Gustafson-Capková
• Stockholm University

2
Issues addressed in research on discourse

• What information is contained in extended
  sequences of utterances that goes beyond the
  meaning of the individual utterances?
• How does the context in which an utterance is
  used affect the meaning of the individual
  utterances, or parts of them?
• (Grosz, 1997)

3
What is discourse?

• A coherent stretch of utterances
• written, spoken, transliterated, acted
• monologue or dialogue
• Diskurs
• Text Tal
• Skrivet Uppläst Agerat Transkriberat
  Talat

4
Coherence

• Coherence cohesion
• Focus
• Anaphora
• Cohesive devices (e.g. discourse markers)
• Discourse relations
• Discourse segments

5
Focus

• Focus can generally be described as the most
  important.
• Sidners process model for focus (1983).
• Focus is signalled by e.g.
• Sentence position (cf Given - New).
• Grammatical role (e.g. CT).
• Syntactic construction (e.g. cleft
  constructions).
• Intonation (focal accent - New).

6
Anaphora

• Anaphoric expressions are said to refer back to a
  concept in focus (e.g. Ariel, 1990, Gundel,
  1995), i.e. focality has been used for finding
  referents for anaphora (e.g. Sidner, 1983).
• The anaphoric expression can also carry
  information (number, gender etc.), and in this
  way constrain the choice of referent.
• However, there are exceptions
• Pelle har köpt en volvo. De är bra bilar
• Jag såg Lisas unge igår. Hon var den sötaste
  lilla bebis jag sett!

7
Abstract object anaphora

• Pelle föll eftersom golvet var så halt. Det var
  det otäckaste jag sett på länge. (Golvet?
  Fallet?)
• If the referent is fallet, which string is
  relevant as antecedent?
• How can we extract the event fallet in a
  general and robust way?
• How is the intonation of the anaphoric pronoun
  affecting the interpretation?
• Pelle föll eftersom golvet var så halt. Det var
  helt nerstänkt med såpa.

8
Discourse markers

• Discourse markers are linguistic expressions that
  are said to indicate
• a discourse relation
• a discourse segment boundary
• However, discourse relations are not dependent on
  discourse markers.

9
Discourse relations

• A discourse relation is the relationship that
  hold between two utterances. The discourse
  relation might be described as an inference.
• The relation that holds in the following example
  can be explanation.
• Pelle föll. Golvet var halt.
• Pelle föll eftersom golvet var halt.

10
Examples

• Pelle föll. Golvet var halt.
• Pelle föll. Golvet var så halt.
• Pelle föll eftersom golvet var halt.
• Pelle föll eftersom golvet var så halt.
• Det var Pelle som föll. Golvet var halt.
• Det var Pelle som föll. Golvet var så halt.
• Det var Pelle som föll eftersom golvet var halt.
• Det var Pelle som föll eftersom golvet var så
  halt.

11
Discourse segments

• What kind of units build a discourse?
• What kind of units are relevant in analysing a
  discourse, with e.g. discourse relations,
  anaphora resolution etc?
• Prosodic phrasing.

12
Spoken discourse

• Focality
• less reduction
• Discourse anaphora
• discourse structure, as signalled with e.g.
  phrasing. Intonation.
• Discourse segmenting
• prosodic phrasing and disfluencies

13
Problems

• Anaphora resolution
• what items are relevant to check by anaphora
  resolution?
• Discourse relations
• How shall we define what inferences that are to
  be drawn?
• Discourse segments
• What units are relevant to work with?

14
Evaluation

• Problem with interrelated discourse dependencies.
• E.g., CT as anaphora resolution algorithm is
  dependent on both utterance segmenting and
  knowledge representation.