Page 157 - DCAP506_ARTIFICIAL_INTELLIGENCE

Unit 11: Natural Language Processing

inference machinery and what knowledge representation to use. Mann and Thompson has Notes
focused more on the meaning of the relations, and has a more surface based approach (i.e.
discourse markers). The treatment of discourse relations has also been integrated in DRT
(Asher, 1993). Asher has used an inventory of discourse relations, much inspired by Mann
& Thompson. Multiple functions by discourse markers have been investigated by e.g.
Hovy (1997), but it is in place to note, that the issue of multiple discourse relations, still is
an open question.
One weakness of the above mentioned approaches to discourse relations is that no
distinction is made concerning on what kind of basis, or from what knowledge sources,
the inferences are drawn. How can we make distinctions between inferences on basis of
what source of knowledge they are computed from? Is the inference based fully on the
semantic content in the utterance, is it based on world/domain knowledge, or is it based
on socio-cultural knowledge. Of course these sources of knowledge might blend, but it is
still the authors feeling, that different kind of knowledge should be separated or at least
marked out in the knowledge representation, as is the attempt by e.g. Sanders et al. (1992).
The inferences, or discourse relations, always holds between parts of the discourse. These
parts are often called discourse segments, but the criteria and strategies in discourse
segmenting is not very well established. However, the discourse segmenting is crucial in
establishing discourse relations, and also in the description of coherence.

Discourse Segments

As mentioned, one way to perform discourse segmenting is on basis of cue phrases. In written
language one can also make use of punctuation and in spoken language the prosodic phrasing.
To establish boundaries implies to make up groupings in between those boundaries.
It is worth to stress that the discourse boundaries do not have to be absolute, i.e. perhaps the
boundary is rather a field than a specific point, and it should also be pointed out that discourse
boundaries might be a function of the grouping. As mentioned earlier, discourse markers have
the double function of dividing parts of the discourse, and at the same time connect those parts
with some discourse or coherence relation. This is quite similar to the role of prosody; Prosody
might also be regarded as a segmenting device, but this segmenting function is double, in that
what defines a boundary also defines a non-boundary. It is an open issue whether we primarily
shall focus on the boundary fields, or on the non boundary-fields. Polanyi (1988) builds discourse
segments on basis of meaning clusters, i.e. she starts on phrase level and merges phrases which
are semantically close to each other into a larger unit. The boundaries then become a result,
when two segments cannot be merged, because they are not semantically close enough. This
kind of discourse segments might thus be described as clusters of meaning. A similar approach
is taken by Hearst (1997), who bases discourse segmenting on term repetition. The parallel to
the prosodic grouping of speech can be made, where the prosodic boundary is not a feature of its
own, but rather a result of the complete prosodic utterance contour. Bruce 1998 stresses this
double function for prosody: Prosody has the function of both boundaries marking and grouping.
However, the textual surface approach with punctuation and cue phrases is more widely
investigated in discourse segmenting, e.g. by Oberlander & Moore (1999), Albritton & Moore
(1999), Flammia (1998). The base for discourse segments are further often claimed to be intention
based (Grosz and Sidner, 1986), or information based (Mann & Thompson, 1988). The difference
lies in whether the segmenting is made on basis of the intentions behind the message, or the
information conveyed in the message. The intentions based approach is quite closely related to
speech act theory and dialogue coding, as described by Carletta et al. (1997). Intentions based
discourse segments and information based discourse segments have traditionally been viewed

LOVELY PROFESSIONAL UNIVERSITY 151