Comprehending Text : Achieving Coherence through a Connectionist Architecture

Comprehending Text: Achieving Coherence through a Connectionist Architecture Samuel W.K. Chan Department of Artificial Intelligence School of Computer Science and Engineering University of New South Wales Sydney NSW 2052 AUSTRALIA Abstract Comprehension is determined by the content of memory. The verbalized concepts reflect the information that is in the focus of attention. Mental models of text must be formed in order to achieve coherence in language understanding. Constructing a mental model requires continual interaction between the text and the reader's linguistic, pragmatic, and world knowledge. In this article, a connectionist framework is proposed for the formation of mental models that occurs during comprehension. A unification and spreading activation model are used to simulate the processes. The system is tested using children's stories. Results attest the validity of the model. Thus, our design contributes to the formation of a durable and functional mental representation of the text information. 1. Introduction Language comprehension is mediated by a complex set of processes that operate on the representations at three main different levels: lexical, syntactic and semantic. Early research on aspects of text representation tended to look at general questions about, for example, the availability of representations of surface form. More recently, comprehension is modeled as construction of mental models of text [1]. Several conclusions about mental representations of discourse have been reached. First, such representations are structurally similar to part of the world rather than to any linguistic structures [2]. "Distilled" meanings of the text and their interrelations are represented directly. This representations are built up as the text is read. Information not explicit in a text may be included in the representations of its content, particularly if it is required to establish links between parts of its content [3]. At any point in a text, the representation constructed up to that point is the context for the interpretation of the next sentence. In particular, it restricts the set of possible referents for anaphoric expressions, and allows such expressions to be interpreted correctly in context [4]. Given this agreement as to the theoretical importance of mental models, however, it is surprising that there is little agreement as to exactly what constitutes a mental model, and there is little research demonstrating the construction of a putative mental model while reading. In this article, our specific goal is to delineate a theoretical connectionist system that generates the cognitive representation for narrative comprehension. In particular, attention has focused on the factors of argument overlapping [5] and situational continuity 39 [6]. Researches in discourse processes have shown these two major factors facilitate readers to construct the mental models. Argument overlapping is analyzed in terms of the number of arguments overlapped. Propositions having the greatest degree of overlapping with other propositions are proved to be the most important. In addition, Gernsbacher [7] has advocated that readers construct mental structures while reading a text and they try to map any incoming information onto the evolving structure. Sentences that maintain a previously established time frame are more likely to be …

Content truncated to 500 words