A Subsymbolic Model of Complex Story Understanding
Peggy Fidelman and Risto Miikkulainen and Ralph Hoffman · Proceedings of the Annual Meeting of the Cognitive Science Society · 2005
A Subsymbolic Model of Complex Story Understanding Peggy Fidelman Risto Miikkulainen Department of Computer Sciences, The University of Texas at Austin {peggy,risto}@cs.utexas.edu Ralph Hoffman Yale-New Haven Psychiatric Hospital [email protected] istic stories. In this section, psychological evidence for scripts will be reviewed, and computational models based on script theory will be outlined. Abstract A computational model of story understanding is presented that is able to process stories consisting of multiple scripts. This model is built from subsymbolic neural networks, but unlike previous such models, it can handle stories of variable structure and length. The model can successfully parse and paraphrase script-based stories that share long sequences of common events, with no confusion between the stories. It also exhibits several aspects of human behavior, including robustness to small changes in the sequence of events and emotion priming effects in response to ambiguous cues. It can therefore serve as a foundation for testing theories of normal and impaired story processing in humans. Script-Based Story Representations Introduction Computational models are a valuable tool for understanding human behavior. They allow investigation of how various theories about cognition may combine to produce observed human function. They can also provide a way of studying im- pairment in a controlled environment, where the behavioral effects of various types of underlying damage can be inves- tigated systematically by lesioning or otherwise disrupting parts of the model. In this paper, a computational model of human story un- derstanding is presented that can learn to read and para- phrase script-based stories of arbitrary length. The model is built from subsymbolic neural networks, which mimic many computational properties of the brain such as distributed rep- resentations and correlation-based learning and performance. This subsymbolic foundation makes it possible to simulate phenomena that arise from these properties, which is difficult to do with symbolic models of story understanding. Unlike previous subsymbolic models, however, it is not restricted to stories consisting of a rigid, fixed-length structure. This flex- ibility allows the processing of more realistic stories, consist- ing of multiple scripts, which in turn allows more meaningful conclusions about human cognition to be drawn. Using a small corpus of hand-designed representative sto- ries, the model is shown to successfully parse and paraphrase stories that share long sequences of common events, with no confusion between the stories. The model also exhibits several aspects of human behavior, including robustness to small changes in the sequence of events and emotion priming effects in response to ambiguous cues. The paper is organized as follows. Section 2 describes related work in script-based story processing, including the DISCERN model on which the current model is based. Sec- tion 3 details the architecture of the model, and Section 4 examines its behavior under various experimental conditions. Section 5 discusses the results of the experiments and possi- ble directions for future work. Background and Related Work Scripts are knowledge structures for stereotypical sequences of events that allow efficient understanding of complex, real- According to script theory (Schank & Abelson, 1977), peo- ple organize knowledge of common routines into stereotypical event sequences. These scripts are made up of sequences of events with open slots, as well as requirements about what can fill those slots. Scripts make interaction efficient by pro- viding everyone involved with a set of expectations about what will take place. For example, most people who have traveled by airplane know that first they must get a board- ing pass, then wait in a security line, then pass through a metal detector, then wait at the gate, and so on. Without such a script, a person would have to put a lot more intel- lectual effort into figuring out what was expected of him at each point. If no one had such a script, airports could hardly serve the function they do. Scripts also serve to make natural language communication efficient. There is no need to recount all the details of an ordinary visit to the dentist, for example; the speaker can just make reference to such a visit, and the listener can fill in the details herself. The hypothesis that humans use such scripts in cognition and language is well supported by experimental evidence. For example, the degree to which events in stories will be remem- bered can be predicted by whether those events are part of such a script (Graesser, Gordon & Sawyer, 1979; Graesser, Woll, Kowalski & Smith, 1980). Similarly, the amount of time it takes for a human to understand a sentence can be predicted by whether it fits into a script (Den Uyl & van Oostendorp, 1980). Because scripts are a particularly well- established theory in psychology, they provide a good foun- dation for a computational model of story processing. Models of Script-Based Story Processing Scripts have been used as a basis for several symbolic models of story processing. The first of these was SAM (Script Ap- plier Mechanism) (Cullingford, 1978), able to handle stories with multiple simultaneously active and interacting scripts. FRUMP (Fast Reading Understanding and Memory Pro- gram) (DeJong, 1979), on the other hand, skimmed newspa- per stories about stereotypical episodes and filled in slots cor- responding to the most important parts of the script. Scripts have been used since then in numerous symbolic systems that aim at understanding natural language stories. Although there has been a lot of work on subsymbolic pro- cessing of sentences in the past two decades (McClelland & Kawamoto, 1986; Jain, 1991; Rohde, 1999; Henderson, 1994; Mayberry, 2004), the approach has been much less successful at the level of stories. Early on, several models were de-