Automating Text Propositionalization: An Assessment of AutoProp
Stephen W. Briner, Philip M. McCarthy, Danielle S. McNamara · eScholarship (California Digital Library) · 2006
Propositions are psychological representations of textual units that capture the overall gist of a sentence or clause (Kintsch, 1988; Kintsch, 1998). Such is the complexity of constructing these propositions that no computational system has yet been able to match human hand-coded examples. The recent advent of major computational projects such as iSTART (McNamara, Levinstein, & Boonthum, 2004) and Coh-Metrix (Graesser et al., 2004) has highlighted the need for an automated tool capable of accurately converting thousands of sentences into propositional units. In this study, we introduce a working prototype of a propositionalization tool, AutoProp, and assess its automated output of propositions against a corpus of published hand-coded propositions. AutoProp, written in Visual Basic, first directs text through the Charniak parser (Charniak, 2000) before allocating the parsed data into propositional units. A final proposition is displayed on AutoProp’s interface, and can be saved to a file or printed upon request. As an example, the sentence The hemoglobin carries the oxygen is represented by Kintsch (1998) as CARRY[HEMOGLOBIN,OXYGEN]. AutoProp separates the sentence into primary elements (pe) and sub-propositional elements (sub prop) rendering the Kintsch sentence above as: carries (the {pe} hemoglobin, {pe} oxygen) sub prop: the ({pe} hemoglobin) sub prop: ({pe} oxygen) For an initial test of the tool’s effectiveness, we constructed a corpus of 29 previously published sentences taken from Kintsch (1998). These sentences were processed through AutoProp to derive the tool’s propositional representations. The types of contrasts (i.e., output differences) between the tool and the Kintsch model were categorized a priori as follows: Type 1: Superficial, easily correctable differences.