| The potential subjectivity of human interpretation is largely ignored in current research on automatic text understanding. Current computational systems are largely based on statistical analysis of text corpora and focus on meaning presumed to exist "in the text", whereas text meaning in computational linguistics in the past and in other disciplines (English, education, and philosophy) is considered to exist "in the text", "in the reader", and "in the writer". This study is an empirical investigation of the subjectivity of readers' perceptions of two related aspects of meaning in text: lexical cohesion and the lexical semantic relations that are its building blocks. Also, the types of lexical semantic relations given by the readers are analyzed. In library and information science, linguistics, computational linguistics, and psychology, lexical semantic relations have largely been studied outside the context of text, and readers' views have not been studied. In this study, participants were given extracts from general-interest articles. They marked groups of words they perceived as related while reading the text, identified related pairs of words within these groups, and stated what the relations were. First, average pair-wise agreement statistics indicate 60% agreement on lexical cohesion as represented by word groups marked by at least 4 readers. For relations between word pairs marked by more than one reader, 70% were similarly described. These results indicate that applications using lexical cohesion and lexical semantic relations for text understanding should consider reader subjectivity and that similar research on the subjectivity of other aspects of text meaning should be conducted. Reader models are suggested to computationally account for interpretations of text meaning that are "in the reader". Thus, a return to considering all three views of text meaning is suggested as computers provide increasingly complex interpretations of text. Second, 62% of the readers' relations were non-classical (not hyponymy, meronymy, synonymy or antonymy). A group of 20 types was found to cover the readers' relation descriptions. These types have potential use in applications that rely on identifying lexical cohesion and/or lexical semantic relations in text. The readers' focus on non-classical relations indicates a need for a similar focus in research and applications. |
