Information Technology Reference
In-Depth Information
2.2.2. Automatic processing
The aspects of NLP linked to automatic understanding [JUR 09] and
automatic generation [REI 00] are obviously involved in MMD, on lexical,
syntactic, semantic and pragmatic aspects. Thus, dictionaries and digital
resources, semantic lexicons or even lexicalized grammars provide content
and methods to store lexical knowledge [MAR 00]. Syntactic formalism,
especially when it has been designed with computational concerns in mind,
allows us to design efficient performance analyzers, which are open to
additional knowledge [ABE 07]. The semantic models, from the conceptual
graphs [SOW 84] and up until the formalisms of formal semantics and
discourse semantics [KAM 93], allow us to combine the contents of words
and utterances [ENJ 05]. Pragmatics, with the resolution of reference in a
purely linguistic or multimodal context [PIN 00], with the identification of the
implicit, such as presuppositions [VAN 02] and that of speech acts [TRA 00],
has also been the focus of various works in NLP which are directly applicable
to MMD.
Other fields of NLP are involved in a more occasional manner, or for a
certain type of dialogue. The resolution of anaphora, which is an aspect of the
resolution of references, is also a field unto itself, with its own algorithms and
assessment campaigns [MIT 02]. An MMD system does not generally need to
implement the most complex algorithms, inasmuch as, in a dialogue, an
anaphora tends to refer to a recent, or even immediately accessible,
antecedent. However, the integration of algorithms that have proved their
worth in their respective fields is obviously beneficial, if the machine's
resources allow it. Another field of NLP, the identification of coreference
chains can be applied to MMD and thus provide additional clarification, for
example on the way the user introduces a new referent and then refers to it;
the system can thus reproduce this behavior at the generation level. The
identification of discourse relations [ASH 03], or the detection of named
entities are also NLP applications useful in MMD, especially for open domain
dialogue.
NLP has not solved all the issues and its limits are visible in MMD. It
is the case of the cover and subtlety of the language, which it cannot record
efficiently and reliably. This is the case when identifying ambiguities, when
managing unknown words, when identifying a non-literal use of the language:
deadpan, irony, sarcasm, exaggeration, rhetorical questions [CLA 96, p. 353].
Search WWH ::




Custom Search