Textual Genre Analysis and Identification - Ambient Intelligence for Scientific Discovery - page 143

Table 4. Multivariate weights in the discriminate functions created using DocuScope's
analysis for Tech Review (Column 1) and Non-Reviews (Column 2). Italicized variables
are highly discriminating and are discussed in this paper.

Tech Review Non-Review

Constant -7.815 -12.826

Thought
First Person: SelfDisclosure

6.005 -7.415

Inner Thinking: SubjectivePercpt 6.813 5.370

Relations
Reasoning: Example 4.196 11.161

Reasoning: StandardNeg 2.279 -1.269

Linear Guidance: Citation 3.631 8.067

Reasoning: Quotation 0.093 4.125

Interacting 2.834 1.707

ShareSocTies: CommonAu 3.524 5.294

Description
PastEvents 2.355 4.684

other texts about PDAs in the ProQuest data base. In our experience, this is
a typical finding. In this case, variables from all three of DocuScope's hierar-
chies (Thought, Relations, and Description) come into play in distinguishing
tech reviews from non-reviews as we discuss in a moment.
To test the validity of DocuScope's model, we then used these discriminate
functions to classify a third set of articles that included a mix of both reviews
and non-reviews. This set consisted of 112 articles published during 1996 that
mentioned personal digital assistants in their titles or abstracts minus ones that
had been included in the original sets. These thus represented the equivalent of
the results from a simple search in the ProQuest database for the year 1996.
To judge how well DocuScope's analysis matched the judgments of a human
rhetorician, we independently classified the same set of texts, deciding whether
the indeed represented reviews or non-review. We then compared our judgments
with DocuScope's judgements.
As Table 5 indicates, DocuScope matched rhetorical judgment most, but
not all, of the time. For those 48 texts that we identified as tech reviews, the

Table 5. Overall, the analyses agreed 82% of the time and disagreed on a total of 20

texts.

DocuScope Judgement
Tech Review Non- Review Totals Misclassified

Rhetorical Tech Review 35 13 48 13 of 48 (27%)

Judgement Non-Review 7 57 64 7of64 (11%)

Totals 42 70 112 20 of 112 (18%)

Correctly Classified 35 of 42 57 of 70 92 of 112

(73%) (89%) (82%)