Semantic Information Processing, edited by Marvin Minsky, is a collection of technical papers on early symbolic computation research. It is not a single-argument monograph or a general textbook; its coherence comes from a shared question: how can a computer store and use the meaning of information—facts, sentences, diagrams, and problems—so that it can answer questions, make inferences, and solve tasks that appear to require understanding? Minsky’s introduction frames this as an engineering problem. A program needs an explicit representation of knowledge, procedures that operate on that representation, and a way to match incoming language or perceptual data against it. The papers that follow are case studies of working programs, each addressed to a different semantic task.
Bertram Raphael’s contribution describes SIR, a program for semantic information retrieval. SIR accepts statements in a restricted English-like input language, stores them as a network of objects and relations, and answers queries by searching that memory and performing simple deductions. The system is meant to retrieve information by meaning rather than by keyword matching: it can relate facts, follow links among entities, and produce answers that were not explicitly stored. Raphael sets out the representation of semantic relations, the input language, and the retrieval procedures, and illustrates the program’s behavior on sample questions.
Daniel Bobrow’s paper describes a natural-language interface to a problem-solving program, commonly known as STUDENT. The system takes algebra word problems stated in English, analyzes them using pattern matching and semantic categories, converts them into equations, and solves the equations. The aim is not general language understanding but a demonstration that restricted natural-language input can be mapped into a formal problem representation. The program relies on knowledge about the problem domain and about English constructions in order to disambiguate wording and set up the mathematics.
Thomas G. Evans’s paper describes a program for solving geometric-analogy questions of the intelligence-test type. Given drawings in the form “A is to B as C is to ?”, the program decomposes the figures into parts and relations, compares the transformation from A to B with candidate transformations from C, and selects the best completion. It is an early attempt to give a machine a structured “perception” of line drawings and to model analogical matching rather than simple feature lookup.
Across these papers, the volume treats meaning as something that can be made concrete in a program: structured symbols, relations, pattern-matching rules, and procedures that search, compare, and transform representations. The chapters include detailed descriptions of data structures and control procedures, along with sample inputs, outputs, and discussions of limitations. The programs are narrow and domain-specific; their restricted vocabularies, hand-built knowledge, and brittle parsing are part of the record rather than hidden. The book is therefore best understood as a set of research reports showing how several early systems attempted semantic processing in language, memory, and problem solving, rather than as a unified theory of meaning.
The volume also includes Ross Quillian’s “Semantic Memory,” which shifts attention from question answering and problem solving to the organization of conceptual memory itself. Quillian proposes representing word meanings as nodes in a network, with labeled links encoding relations among concepts and their properties. Semantic comparison and retrieval are modeled as search over this network: activation spreads from concepts, associations are followed, and meanings are related by finding paths or intersections between them. The paper is concerned with how definitions and facts might be stored in a shared structure rather than as isolated entries, and with the procedures that would traverse such a structure to compare meanings or answer questions. It supplies an explicit semantic-network proposal and discusses the difficulties of encoding concepts and constraining search over associative links. In the context of the collection, it broadens the treatment of semantic processing to include the memory organization that language use and inference presuppose, alongside the book’s programs for retrieval, language-to-equation mapping, and analogical matching.
