This dissertation studies indirect anaphora (IA) resolution in English narrative discourses from a cognitive perspective. Though the study of IA in English language was initiated long time ago, compared with the study of direct anaphora (DA), not so much relevant literature can be found about this research. Among the limited literature, people can find three main theoretical frameworks that work more or less effectively in IA resolution. They are: the Topic/Focus Model, the Scenario Model and the Relevance Model. The Topic/Focus Model is in the main a syntactic model. The Scenario Model is one that is vigorously backed up by psychological experiments. And the Relevance Model is the newest and rather impressive model. However, each theoretical account itself has some evident defects that can hamper the tracking of expected antecedent in more complicated cases, and the application of each theoretical account in previous literature also contains some problems, such as overemphasis on stereotypic roles and negligence of on-line instances. As a result, these theoretical accounts leave much room for improvement or modification. Mainly for the above reasons, this dissertation proposes a cognitive approach to IA resolution in on-line processing of English narrative discourses. The basic hypothesis of this dissertation is that IA resolution is fundamentally a process of the association of mental entities, a process of restoring the previously-established relations. So, the basic approach to IA resolution must be associated with one’s mental activities. In doing so, this dissertation has no intention of belittling the importance of surface linguistic cues. This dissertation holds that linguistic expressions and contextual cues are triggers of mental representations, scenarios and relations. Both the contextual cues on the surface linguistic level and the mental activities on the psychological level are important, but the latter is of crucial importance and is the focus for detailed investigation.The data of this dissertation covers 67 English short stories with the number of words totaling approximately 770,000. From this pool of data have been drawn all together 483 IA instances, which are classified and analyzed in detail. This dissertation finds that there are a lot of striking features in these on-line instances which are different from many of the earlier researches, such as the diversified NP1’s forms, the shift of NP1’s syntactic position to the front of NP2. IA resolution is not simply a matter of the association between antecedent and anaphor; a variety of contextual cues are involved in the tracking process. This dissertation finds from the data five major contextual cues, which are: the semantic relation with NP1, pragmatic strengthening or weakening effect, syntactic position, recency of mention and parallel structure. Our analysis of the data shows that in reference tracking there is frequently one or two factors that play decisive roles; all the other factors exert only trivial influence on the tracking process, and such effects can be ignored. Based on the on-line instances that are collected, this dissertation proposes a set of Basic Tracking Principles, which can resolve most of the stereotype-driven IA instances. Although there are some obvious differences between stereotype-driven and non-stereotype-driven IA instances, they share an extremely important property, i.e. relatively privileged associative strength, which is the ultimate cause for the selection of the expected antecedent among two or more possible antecedents. We finally propose a Unified Tracking Principles that can resolve both stereotype-driven and non-stereotype-driven IA instances. 内容来自dedecms