Question-Answering by a Semantic Network of Parallel Automata

Human semantic memory is modeled as a network with a finite automation embedded at each node. The nodes represent concepts in the memory, and every arc bears a label denoting the binary relation between the two concepts that it joins. The process of question-answering is formulated as a mathematical problem: Given a finite sequence of labels, tlnd a path in memory between two given nodes whose arcs bear that sequence of labels. It is shown that the network of automata can determine the existence of such a path using only local computation, meaning that each automaton communicates only with its immediate neighbors in the network. Furthermore, any node-concept along the solution path can be retrieved. The question-answering algorithm is then extended to incorporate simple inferences based on the equivalence of certain sequences of relational labels. In this case, it is shown that the network of automata will find the shortest inferable solution path, if one exists. Application of these results to a semantic corpus is illustrated. The two semantic topics of negation and quantification receive special treatment. Careful study is made of the network structure required to encode information relating to those topics and of the question-answering procedures required to extract this information. The notions of a negated relation and a negated question are introduced, and a negation-sensitive path-searching algorithm is developed that provides for strong denials of queries. For sentences involving universal and existential quantifiers, it is shown how the terminal can translate a first-order language question into a sequence of network queries. In both areas, the network model makes reaction-time predictions that are supported by several experimental findings. Extensions of the model that would permit the encoding and retrieval of propositional information are mentioned.