Textual Signatures: Identifying Text-Types Using Latent Semantic Analysis to Measure the Cohesion of Text Structures - Natural Language Processing and Text Mining - page 112

Fig. 7.2. Average length in termsof number ofwords in eachtext section.
7.9 Experiment 2
IfLSAcompares relativesimilarities between sections, thenitis reasonable to as-
sume that comparingsimilarly themed papers wouldproduce signatures similar to
those observedinExperiment 1. Because less ofarelationship isexpectedfrom
papers with overlapping themes ascomparedto those fromthesame paper, we pre-
dictedarelative decline in the LSA comparison scores. We also expectedareduced
relationship betweenthe abstract and resultssectionbecause theyare fromdifferent
studies.
To testthis prediction, wecomposedanentirelynew corpusof 20 similar articles:
all themedas working memory and intelligence . We then extractedthe abstractsof
these texts fromthe bodies and randomlyreassignedthe abstractto adifferent
articles parts. As in Experiment 1, weconductedarepeatedmeasures ANOVA on
the LSA cosines including the within-text factorsofAI( M=.439, SD=.136 ), AM
( M =.289, SD =.110), AR ( M =.304, SD =.144), and AD ( M =.454, SD =.162). There
was a main effect of comparisontype, F (3,19)= 16.548, MSE =.009, p< .001. Pairwise
contrasts(seeTable 7.3) indicated that all differences were reliable except between
AM and AD and betweenAMandAR.
Asshown in Figure 7.1, the pattern of cosines issimilar to that of Experiment
1, with reduced scores overall comparedto Experiment 1, and a reductioninthe
relationship betweenthe abstracts and resultssection. These results allowus to
predict that LSA can produce prototypicalsignatures that are able to differentiate
between sections fromthesame articles, and those articles that are merely similar
in theme.