On Removing Ambiguity in Text Understanding

This paper discusses how to remove a kind of ambiguity in a text understanding system based on simulation. The system simulates some events mentioned in text on a world model and observes the behavior of the model during the simulation. Through these processes, it can recognize the other events mentioned implicitly. However, in case the system infers plural number of possible world, it can't decide which one is consistent with the context. We deal with such ambiguities. The ambiguities, in some cases, can be removed by considering contextual information. To remove the ambiguities in the simulation, we define three heuristics based on the characters of the explanatory descriptions and propose an algorithm of "looking ahead". We implement ,ani experimental system. Accqrding to the ,algorithm, the system finds out the supplementary descriptions in the following sentences and removes the ambiguities by using the contents of the supplementary descriptions. In this paper, we discuss a method of removing ambiguities which appear in the process of text understanding based on simulation. We have been studying a text understanding system paying attention to the importance of an imagerial world model. In [Itoh,92, Itoh,95], we showed that the ability to simulate the matters described by sentences on the imagerial world model is one of the basic abilities to understand texts. We also proposed a method of implementing an imagerial world model, a method of simulating on the model and a method of observing the model to extract propositional expressions. We introduced an experimental ,system in order to verify importance of the imagerial world model and validity of each method. In order to clarify the point of issue, we- have done the work under the following constraints. 1. We restrict our target texts to those explaining mechanical movements of machines in textbooks for junior high school students or encyclopedias for naive persons. The reasons are that we need to treat a relatively narrow domain in order to implement a world model imagerially in a simple way, and imagerial information seems to be quite important in the domain of mechanical movements. 2. We restrict the machines to those which are composed of solid parts and are illustrated with two dimensional figures. It also makes . the implementation of the model easy. 3. We deal with only the sentences which explain states or movements of a whole machine or its parts. Though there are some sentences explaining a pressure of gaseous fuel, a flow of fluid, or etc., we neglect them. If we deal with such sentences, we should represent &pressure or a flow of fluid as well as positions and shapes of machine. In other words, the imagerial world model should hold information on multiple attributes. It makes the imagerial model complex.