Page 155 - DCAP506_ARTIFICIAL_INTELLIGENCE

Unit 11: Natural Language Processing

is “the most important element”. An important question, which complicates the picture, is Notes
then: important for whom/what? I.e. important to e.g. the speaker, the listener, the global
message or the local utterance? Thus, focus is a concept simple to understand, but it is
hopelessly difficult to pin down exactly. In this paper the term is used in the meaning
salience or focus of attention. If it is used in some other meaning, this is indicated in the
text. For a survey of the different uses of the term focus, see Gundel (1995). The feature
focus is connected to cognitive processing, memory representation and linguistic
representation, in that a more activated/Given/accessible concept is represented by a
more reduced linguistic sign (Ariel, 1990, Gundel, Hedberg & Zacharski, 1993). Of course
focus is also connected to the syntax of the utterance and the semantics (e.g. Sidner, 1983)
as well as the prosody (Bruce, 1998, van Donzel, 1999). Van Donzel made extensive
investigations on basis of Princes (1983) Given – New taxonomy, and her findings confirms
the picture that Given elements are also acoustically more reduced than New elements.
Thus, the intonation plays a great part for the information structure in spoken language,
however, it is also possible to do without it, as in written language. Ability to predict
focus should improve the assignment of prosodic features in speech synthesis, however,
the task seems to be very difficult, (if not impossible) Bolinger (1972), and the result in
synthesis are also today limited. Focality is further an often used feature in anaphora
resolution.
2. Anaphora Resolution: Very closely connected to the research on focus and also to the
research on discourse coherence is the research on anaphora resolution. This because
anaphora is affected by focality and further context dependent. This means that anaphoric
expressions are representing referents in focus, and they are at the same time establishing
links back to the preceding context (the antecedent).
Traditionally, the concept of anaphora was limited to co-referent NPs, this means that the
only antecedents we have to keep available in the discourse record are the NP objects.
However, this approach becomes to narrow if one want to give account for more complex,
or implicit kind of anaphora, such as situation anaphora, abstract object anaphora or
associative anaphora (e.g. Webber, 1990, Fraurud, 1992, Asher, 1993, Dahl & Hellman,
1995), or connected phenomena treated as anaphora such as presuppositions (van der
Sandt, 1992) and referent coercion (e.g. Dahl & Hellman, 1995).
It is clear, that to refer back to those more abstract or implicit objects, it is not enough to
just keep NP objects available in the discourse record. Rather it is needed a discourse
record with information about e.g. situations and associative connections (semantically
related items). It is also worth noticing that while co-referent NP anaphora has a clearly
anaphoric identity, the abstract object anaphora and the associative anaphora might be
more easily understood as discourse relations, e.g. rhetorical relations (Mann & Thompson,
1988) or coherence relations (Hobbs, 1995). There is a large body of work on automatic
anaphora resolution, however, often it is delimited to resolution of NP anaphora. Sidner
(1983) makes explicitly use of the concept of focus, for resolving anaphora, i.e. her anaphora
resolution algorithm is in the first step predicting focus, and in the second step choosing
the antecedent on basis of the focus ranking. The use of Centering Theory (Grosz, Joshi &
Weinstein, 1995) as an algorithm for anaphora resolution works in the same way, i.e. (a)
predict focus, (b) choose antecedents from the ranked elements. Eckert and Strube (1998)
has used the Centering theory algorithm, extended with speech act tagging, to develop an
algorithm for resolution of abstract object anaphora in dialogue. Also Harabagiu (1999)
have done work on automatic co-reference resolution.
To investigate anaphora might be viewed as a way to investigate the “symptom” of
coherence in discourse.

LOVELY PROFESSIONAL UNIVERSITY 149