Page 156 - DCAP506_ARTIFICIAL_INTELLIGENCE

Artificial Intelligence

Notes 3. Cohesion and Coherence: Coherence is what makes a collections of sentences/utterances a
discourse, but what exactly is it? Coherence might be defined as implicit relations between
different parts of the discourse. Coherence is closely connected to the concept of cohesion,
which means explicit markers of relations between different parts of the discourse.

Example: Coherence is in many cases established with cohesive devices, examples of
such cohesive devices is e.g. discourse markers or anaphoric pronouns. E.g. Harabagiu (1999)
stresses the importance of cohesive devices for coherence in discourse.
Discourse markers, or cue phrases, are said to indicate discourse relations, but they are
also said to have the function of marking out discourse boundaries, i.e. to perform discourse
segmenting.

Example: Some discourse markers are: “but”, “anyway”, “okay” etc.
Research aiming to give account for coherence in discourse has been conducted by e.g.
Grosz & Sidner (1986), who have developed a discourse theory, which gives account for
coherence on a global level of discourse, but with a limited inventory of discourse relations.
The discourse relations by Grozs & Sidner (1986) are rather solely structure indicating
(i.e. mother-daughter relation or sister relation).

Did u know? Discourse markers are primary seen as segmenting devices, and the specific
relational information is not stressed.

Grosz, Joshi & Weinstein (1995) have with CT developed a way to give account for the
degree of coherence in a discourse. Their approach to coherence is based on anaphoric
relations, and does not pose any extra meaning in the relations between utterances.
The CT has been suggested many modifications, e.g. by 1998, Strube & Hahn, (1998, change
of ranking criteria), Walker (1997, change of discourse segmenting), Passonneau, (1993,
adding a more hierarchical structure), Beaver, (2000, rewriting CT in Optimality theory).
Also evaluations have been made, e.g. by Byron & Stent (1998), Poesio et al. (2000), Tetrault,
(forthc).
4. Discourse Relations: Discourse markers are often seen as establishing a relation between
parts of the discourse, e.g. the discourse marker “but” can establish the relation contrast
between two utterances or discourse segments. In this case we can see the relation as
marked on the surface level. The relation “contrast” can furthermore be signaled
prosodically.
The role of the meaning by discourse relations might be more stressed than by e.g. Grosz
& Sidner (1986). In this case the discourse relations can be described as inferences drawn on
basis of what is said in the discourse. The inferences thus enriches the discourse meaning,
and might also function as a glue between utterances/segments, and the inferences then
contributes to the meaning of the message, and not just to the structure. The inferences can
be drawn on basis of adjacency, but are still mainly semantic to their nature, i.e. they are
not dependent on the adjacency in the same way as the surface signaled discourse relations
are. In the case of inferences, the inference machinery, as well as the knowledge
representation is of course crucial.

There have been many attempts to give account for coherence in discourse, in terms of
discourse relations, e.g. by Hobbs (1985) and by Mann & Thompson (1988). Specially
Hobbs have been interested in how to compute the discourse relations, on basis of what

150 LOVELY PROFESSIONAL UNIVERSITY