Mining Diagnostic Text Reports by Learning to Annotate Knowledge Roles - Natural Language Processing and Text Mining - page 46

of measurement and test results, overall assessment and recommendations, as well as
several attachments with graphical plots of numerical measurements or photographs
of damaged parts.
From a diagnostic point of view, the most important information is found in the
evaluations of the measurements and tests performed. As a demonstration, consider
the two excerpts in Figure 4.1 (originating from English documents for non-German
speaking customers).

At 1 . 9 U N (= 30 kV ), an insulation breakdown
occurred on the upper bar of the slot N ◦ 18,
at the slot exit on the NDE side. The break-
down indicates that the bar insulation is se-
riously weakened. This may be caused by in-
tense discharges due to a malfunction of the
slot anti-corona protection.

The measured bypass currents are in a rel-
atively high range indicating a certain sur-
face conductivity. This is due to the fact that
the motor was stored in cold area before it
was moved to the high voltage laboratory
where the temperature and humidity was much
higher so that a certain degree of condensation
could occur on the surface of the winding.

Fig. 4.1. Excerpts from two evaluations of isolation current measurements.

As it is often the case with diagnosis, while the quantities that are measured
or the components that are inspected are the same, the findings depend on a series
of contextual factors, and the reasons for these findings could be quite unique (as
the examples of Figure 4.1 demonstrate). Usually, human experts need many years
of field experience to gain a degree of expertise that allows them to handle any
situation. The goal of our project is to mine the text documents for relevant pieces
of knowledge acquired during diagnostic problem solving situations.

4.2.2 Domain Concepts

In some text mining applications, such as text categorization or information retrieval,
the goal is often to discover terms specific to the domain that could be used as
indices for organizing or retrieving information. Indeed, the excerpts of Figure 4.1
contain several of such domain-specific terms: insulation , discharge , slot anti-corona
protection , conductivity ,or winding . Still, using these terms as indices or keywords
for representing the documents does not contribute to the purpose of our intended
application, which is to find knowledge that supports diagnostic problem solving.
To exemplify, consider the sentences in Figure 4.2:

1) The calculated insulating resistance values lay in the safe operating area.
2) Compared to the last examination, lower values for the insulating resistance
were ascertained, due to dirtiness at the surface.

Fig. 4.2. Two sentences with the same domain concept shown in boldface.

In both sentences, the domain concept insulating resistance is found, but from a
diagnostic point of view only the second sentence is interesting, because it describes