VERTICAL APPLICATIONS:
How "Sequence Packages" Can Aid Language
Understanding
Method Used to Analyze Doctor-Patient Interviews
Can Be Applied to Customer Service Systems
By Amy Neustein
Critical patient history is often buried in the convoluted, ambiguous utterances
that occur in the doctor-patient interview.
When I was engaged in the analysis of doctor-patient discourse in the 1980s, I
proposed the development of an expert system for identifying the important patient
information that can so easily escape notice.
The method, which I referred to as "sequence packages," was based on
lexical components and paralinguistic features (prosody) as they occur in natural
language.
When a doctor-patient interview is examined, questions, answers, invitations,
complaints, accusations, disagreements, and requests can be seen to appear within tightly
organized utterance sequences displaying the highly systematic and strongly organized
nature of talk.
By the process of identifying specific sequence packages located in the
patients conversation, one is then able to uncover historical data that may
otherwise be unobtainable by virtue of their masked presentation.
The sequence package approach of understanding medical history dialogue can be
applied to a wide spectrum of applications. Just as a patient encounters difficulty in
clearly articulating symptom descriptions, a consumer may also display problems in
articulating product descriptions.
The frustrated consumer, who has a rough idea of a desired product but who can
not effectively communicate the product features to the customer service operator (or to
an automated system) may likewise benefit from a sequence package analysis of his/her
utterances.
Speech recognition companies are now entering a phase where they can offer
natural language dialogues where a user carries on a conversation with a computer in the
same fashion that one would talk with another human being.
In a service-oriented society, consumers often use phone related services to
report problems and obtain remedial assistance. Callers also often need emphatic support
as a way of validating the legitimacy of their complaints.
Studies show that human operators often misread the caller. Human operators can
easily offer too much remedial support, when the callers are seeking some form of
validation. Or they may proffer empathy when the caller required remedial assistance.
If a computer were designed to identify the callers sequence packages
which were consistent with either a quest for remedial support or a search for sympathy
then the computers ability to respond to the callers needs would be enhanced.
As automatic speech recognition systems become increasingly more human-like, a
system that genuinely understands what the callers are requesting would exponentially
increase user satisfaction.
Idioms
Now the question remains as to how "intelligent" are operating systems
required to be as a precondition to engage in natural language dialogue with the user.
This requires the incorporation of some type of artificial intelligence.
But what happens when callers use idiomatic expressions, colloquialisms,
shibboleths, and other unintelligible (to the system) discourse features? No system can
completely cover the variety of utterance components.
If one assumes the starting point of constructing algorithms to correspond to
sequence packages, a computer may be able to identify specific sequence packages in the
naturally occurring dialogue with the user.
Idiomatic expressions are often packaged within utterance sequences that contain
hyperbolic adjectival descriptors ("I did absolutely everything I could." Or
"Ive spent every single waking moment focused on this problem!").
Studies have found idioms to often generate these sorts of hyperbolic
descriptors. No longer is the idiom heard as "unintelligible" once the system
contains the algorithm showing idiomatic expressions as integral to hyperbolic
descriptors. Once the system has acquired context-dependent meaning the system can more
easily engage in the appropriate natural dialogue.
In summary, a sequence package approach to natural language understanding can
demystify the complex cognitive processes of natural language processing and permit the
construction of a system that does not require sophisticated and exhaustive artificial
intelligence to engage a user in a natural language dialogue.
Amy Neustein is the president of Linguistic Technology Systems, Inc., 135
East 54th St., Suite 7J, New York, NY, 10022, and can be reached at
212-605-9926 or lingtec@banet.net.