Participant Reference in Narrative Discourse: A Comparison of Three Methodologies
In recent years several linguists have developed quantitative methods for analyzing and describing how speakers of languages refer to activated participants in narrative discourse. One of the first of these was Talmy Givon, whose method (referred to by some as the Recency/Distance method) measured three factors in participant reference: referential distance, potential interference, and persistence. A few years later, Russell Tomlin proposed an alternative model, which he labeled the Episode/Paragraph method. He felt that Givon's method fails to account for the fact that the thematic paragraphs or "episodes" found in narrative texts largely determine the amount of coding material used. His method assesses the point within an episode at which a referring expression occurs.
Stephen H. Levinsohn developed a third strategy for analyzing referential systems in narrative texts. Informally called the Default/Marked method, he suggests categorizing subject and non-subject nominals into one of several contexts, then determining the default level of encoding for each context. Deviant tokens are then studied to determine the motivation for the deviation.
With such widely different approaches, the question arises as to whether one seems to more accurately and thoroughly describe participant reference than do the others.
A single, lengthy narrative text from the Sio language of Papua New Guinea was selected for the analysis. The first step was to prepare a chart showing the various constituents of each clause in the text. Tabulations were made regarding the participants in the text, including the clause position in which they were introduced and the amount of coding material used. Then, the Recency/Distance method was applied, and values were calculated for each of the three major factors. The next step was to identify the episode boundaries within the text, and apply the xiii Episode/Paragraph method to determine how many over- or under-coded tokens could be explained by their position within episodes.
Then, the Default/Marked method was applied to the text. Default encoding levels were obtained for each subject and non-subject context, using detailed analyses of each participant in every clause. Once this was complete, those tokens that had an amount of coding material other than the default level for that context were analyzed, to determine possible reasons for this marked encoding. Finally, the results obtained using each of the three methods were assessed.
It was determined that the Default/Marked method provides the most thorough analysis of participant reference. The Recency/Distance method demonstrated the relationship between referential distance and the amount of coding material, but produced less convincing results for the other two factors, and failed to account for certain important discourse features and structure. The Episode/Paragraph method proved to be even less effective, as the number of over- and under-coded tokens explained by the position in arbitrarily pre-determined episodes was quite small.
In contrast, the Default/Marked method accounted for all aspects of the referential system in Sio. Default levels for each of the contexts were easily obtained, and the motivation for nondefault (marked) encoding was as predicted by the theory. Thus it serves as a comprehensive and flexible method, one which undoubtedly can be effectively applied to narrative discourse in any language. [ In the interest of making this work available without further delay, we are posting it as it was accepted by the institution that granted the degree without further peer review.]