Assessing the Effect of Text Type on the Choice of Linguistic Mechanisms in Scientific Publications

In this paper, we report a qualitative and quantitative evaluation of a hand-crafted set of discourse features and their interaction with different text types. To be more specific, we compared two distinct text types—scientific abstracts and their accompanying full texts—in terms of linguistic properties, which include, among others, sentence length, coreference information, noun density, self-mentions, noun phrase count, and noun phrase complexity. Our findings suggest that abstracts and full texts differ in three mechanisms which are size and purpose bound. In abstracts, nouns tend to be more densely distributed, which indicates that there is a smaller distance between noun occurrences to be observed because of the compact size of abstracts. Furthermore, in abstracts we find a higher frequency of personal and possessive pronouns which authors use to make references to themselves. In contrast, in full texts we observe a higher frequency of noun phrases. These findings are our first attempt to identify text type motivated linguistic features that can help us draw clearer text type boundaries. These features could be used as parameters during the construction of systems for writing evaluation that could assist both tutors and students in text analysis, or as guides in linguistically-controllable neural text generation systems.